earthquake prediction using machine learning research paper

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• My Account Login
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Open access
• Published: 06 August 2021

Machine learning and earthquake forecasting—next steps

• Gregory C. Beroza 1 ,
• Margarita Segou 2 &
• S. Mostafa Mousavi 1  

Nature Communications volume  12 , Article number:  4761 ( 2021 ) Cite this article

19k Accesses

59 Citations

48 Altmetric

Metrics details

• Natural hazards

A new generation of earthquake catalogs developed through supervised machine-learning illuminates earthquake activity with unprecedented detail. Application of unsupervised machine learning to analyze the more complete expression of seismicity in these catalogs may be the fastest route to improving earthquake forecasting.

The past 5 years have seen a rapidly accelerating effort in applying machine learning to seismological problems. The serial components of earthquake monitoring workflows include: detection, arrival time measurement, phase association, location, and characterization. All of these tasks have seen rapid progress due to effective implementation of machine-learning approaches. They have proven opportune targets for machine learning in seismology mainly due to the large, labeled data sets, which are often publicly available, and that were constructed through decades of dedicated work by skilled analysts. These are the essential ingredient for building complex supervised models. Progress has been realized in research mode to analyze the details of seismicity well after the earthquakes being studied have occurred, and machine-learning techniques are poised to be implemented in operational mode for real-time monitoring. We will soon have a next generation of earthquake catalogs that contain much more information. How much more? These more complete catalogs typically feature at least a factor of ten more earthquakes (Fig.  1 ) and provide a higher-resolution picture of seismically active faults.

a Real-time catalog, available at http://cnt.rm.ingv.it/ and ( b ) machine-learning catalog 16 are shown for event magnitudes above their respective magnitude of completeness 12 , 16 Mc = 2.2 and Mc = 0.5.

This next generation of earthquake catalogs will not be the single, static objects seismologists are accustomed to working with. For example, less than 2 years after the 2019 Ridgecrest, California earthquake sequence there already exist four next-generation catalogs, each of which were developed with different enhanced detection techniques. Now, and in the future, this will be the norm, and earthquake catalogs will be updated and improved—potentially dramatically—with time. Second-generation deep learning models 1 that are specifically designed based on earthquake signal characteristics and that mimic the manual processing by analysts, can lead to performance increases beyond those offered by earlier models that adapted neural network architectures from other fields. Those interested in using earthquake catalogs for forecasting can anticipate a shifting landscape with continuing improvements.

While these improvements are impressive, the value of the extra information they provide is less clear. What will we learn about earthquake behavior from these deeper catalogs and how might it improve the prospects for the stubbornly difficult problem of earthquake forecasting?

Short-term deterministic earthquake prediction remains elusive and is perhaps impossible; however, probabilistic earthquake forecasting is another matter. It remains the subject of focused and sustained attention and it informs earthquake hazard characterization 2 and thus both policy and earthquake risk reduction. A key assumption is that what we learn from the newly uncovered small earthquakes in AI-based catalogs, will inform earthquake forecasting for events of all magnitudes. The observed scale invariance of earthquake behavior suggests this is a reasonable expectation.

Empirical seismological relationships have played a key role in the development of earthquake forecasting. These include Omori’s law 3 that describes the temporal decay of aftershock rate, the magnitude-frequency distribution, with the b-value describing the relative numbers of small vs. large earthquakes 4 , and the Epidemic Type Aftershock Sequence (ETAS) model 5 in which earthquakes are treated as a self-exciting process governed by Omori’s law for their frequency of occurrence and Gutenberg–Richter statistics for their magnitude. These empirical laws continue to prove their utility. Just in the past few years, the time dependence of the b-value has been used to try to anticipate the likelihood of large earthquakes during an ongoing earthquake sequence 6 and the ETAS model has been improved to better anticipate future large events 7 . So it appears that there is value in applying these longstanding relationships to improved earthquake catalogs, but our opinion is that much more needs to be done.

The relationships cited above date from 127, 77, and 33 years ago. The oldest of them, Omori’s Law, was developed based on felt reports without the benefit of instrumental measurements. We suggest that a fresh approach using more powerful techniques is warranted. Earthquake catalogs are complex, high-dimensional objects and as Fig.  1 makes clear, that is even more true for the deeper catalogs that are being developed through machine learning. Their high dimensionality makes them challenging for seismologists to explore, and the conventional approaches noted above seem unlikely to be taking advantage of the wealth of new information available in the new generation of deeper catalogs. We suggest that, having first enabled the development of these catalogs, the statistical-learning techniques of data science are now poised to play an important role in uncovering new relationships within them. The obvious next step is to apply the techniques of machine learning in discovery mode 8 to discern new relationships encoded in the seismicity.

There are tantalizing indications that such an approach may lead to new insights. In double-direct-shear experiments, background signals that were thought to be uninformative random noise have instead been shown to encode information on the state of friction and the eventual time of failure of faults in a laboratory setting 9 . Well-controlled laboratory analogs to faults lack the geologic complexity of the Earth, yet, weak natural background vibrations of a similar sort, that again were thought to be random noise, have been shown to embody information that can be used to predict the onset time of slow slip events in the Cascadia subduction zone 10 . Finally, unsupervised deep learning, in which algorithms are used to discern patterns in data without the benefit of prior labels, applied to seismic waveform data uncovered precursory signals preceding the large and damaging 2017 landslide and tsunami in Greenland 11 .

These examples are compelling but come with the caveat that they are not representative of the typical fast rupture velocity earthquakes on tectonic faults that are of societal concern. For such earthquakes, however, there are also indications from state-of-the-art forecasting approaches that next-generation earthquake catalogs may contain information that will lead to progress. Physics-based forecasting models, which account for changes in the Coulomb failure stress due to antecedent earthquakes that favor the occurrence of subsequent earthquakes, have shown increasing skill such that they are competitive with, and are beginning to outperform, statistical models. Coulomb failure models benefit particularly from deeper catalogs because they include many more small magnitude earthquakes. These small earthquakes add predictive power through their secondary triggering effects tracking the evolution of the fine-scale stress field that ultimately controls earthquake nucleation in foreshock and aftershock sequences. They can also be used to define the emerging active structures that comprise fault networks and by doing so clarify the relevant components of stress that would act to trigger earthquakes 12 . Secondary triggering and background stress heterogeneity were shown to improve stress triggering models 13 but were most effective when they incorporated near‐real‐time aftershock data from the sequence as it unfolded 14 . We note that there is no reason why more complete earthquake catalogs, developed with pre-trained neural network models, cannot be created in real time as an earthquake sequence unfolds. Finally, despite the disappointing history of the search for precursors, due diligence requires that seismologists consider the pursuit of signals that might be precursory.

We conclude that it is now possible to image the activity on active fault systems with unprecedented spatial resolution. This will enable experimentation with familiar hypotheses and enable the formulation of new hypotheses. It seems certain that the underlying processes that drive earthquake occurrence are encoded in this next generation of earthquake catalogs, but we may not find them unless we put new effort into searching for them. Unsupervised learning methods 15 are particularly well-suited tool for that effort.

Mousavi, S. M. et al. Earthquake Transformer - an attentive deep-learning model for simultaneous earthquake detection and phase picking.  Nat. Commun. 11 , 3952, https://doi.org/10.1038/s41467-020-17591-w (2020).

Field, E. H. et al. A synoptic view of the third Uniform California Earthquake Rupture Forecast model.  Seismological Res. Lett. 88 , 1259–1267, https://doi.org/10.1785/0220170045 (2017).

Omori, F. J. On the aftershocks of earthquakes.  Coll. Sci. Imp. Univ. Tokyo 7 , 111–200 (1894).

Google Scholar  

Gutenberg, B. & Richter, C. F. Frequency of earthquakes in California.  Bull. Seismol. Soc. Am. 34 , 185–188 (1944).

Article   Google Scholar  

Ogata, Y. Statistical models for earthquake occurrences and residual analysis for point processes.  J. Am. Stat. Assoc. 83 , 9–27 (1988).

Gulia, L. & Wiemer, S. Real-time discrimination of earthquake foreshocks and aftershocks.  Nature 574 , 193–199, https://doi.org/10.1038/s41586-019-1606-4 (2019).

Shcherbakov, R. et al. Forecasting the magnitude of the largest expected earthquake. Nat. Commun. 10 , 4051, https://doi.org/10.1038/s41467-019-11958-4 (2019).

Bergen, K. J. et al.  Machine learning for data-driven discovery in solid earth geoscience.  Science 363 , eaau0323, https://doi.org/10.1126/science.aau0323 (2019).

Article   CAS   PubMed   Google Scholar  

Rouet-Leduc, B. et al. Machine learning predicts laboratory earthquakes.  Geophys. Res. Lett . 44 , 9276–9282 https://doi.org/10.1002/2017GL074677 (2017).

Hulbert, C. et al. An exponential build-up in seismic energy suggests a months-long nucleation of slow slip in Cascadia.  Nat. Commun. 11 , 4139, https://doi.org/10.1038/s41467-020-17754-9 (2020).

Article   ADS   CAS   PubMed   PubMed Central   Google Scholar  

Seydoux, L. et al. Clustering earthquake signals and background noises in continuous seismic data with unsupervised deep learning.  Nat. Commun. 11 , 3972, https://doi.org/10.1038/s41467-020-17841-x (2020).

Mancini, S. et al. Improving physics-based aftershock forecasts during the 2016-2017 Central Italy earthquake cascade.  J. Geophys. Res. Solid Earth 124 , https://doi.org/10.1029/2019JB017874 (2019).

Segou, M., & Parsons, T. A new technique to calculate earthquake stress transfer and to probe the physics of aftershocks.  Bull. Seismol. Soc. Am . 110 , 863–873,  https://doi.org/10.1785/0120190033 (2020).

Mancini, S. et al. The predictive skills of elastic Coulomb rate-and-state aftershock forecasts during the 2019 Ridgecrest, California, earthquake sequence.  Bull. Seismol. Soc. Am. 110 , 1736–1751, https://doi.org/10.1785/0120200028 (2020).

Mousavi, S. M., et al. Unsupervised clustering of seismic signals using deep convolutional autoencoders.  IEEE Trans. Geosci. Remote Sens. Lett. 16 , 1693–1697, https://doi.org/10.1109/LGRS.2019.2909218 (2019).

Article   ADS   Google Scholar  

Tan, Y. J. et al. Machine-learning-based high-resolution catalog reveals how complex fault structures were activated during the 2016-2017 Central Italy sequence.  Seismic Rec. 1 , 11–19, https://doi.org/10.1785/0320210001 (2021).

Download references

Acknowledgements

This work is supported by the NERC-NSFGEO 13 funded project The Central Apennines Earthquake Sequence under a New Microscrope (NE/R0000794/1). G.C.B. was supported by Department of Energy (Basic Energy Sciences; Award DE-SC0020445). Thanks to Dr. Simone Mancini for preparing the figure.

Author information

Authors and affiliations.

Department of Geophysics, Stanford University, Stanford, CA, USA

Gregory C. Beroza & S. Mostafa Mousavi

British Geological Survey, Research Avenue South, Lyell Centre, Edinburgh, UK

Margarita Segou

You can also search for this author in PubMed   Google Scholar

Contributions

All authors contributed to the development of this comment, and all were involved in revising it.

Corresponding author

Correspondence to Gregory C. Beroza .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Beroza, G.C., Segou, M. & Mostafa Mousavi, S. Machine learning and earthquake forecasting—next steps. Nat Commun 12 , 4761 (2021). https://doi.org/10.1038/s41467-021-24952-6

Download citation

Received : 28 May 2021

Accepted : 12 July 2021

Published : 06 August 2021

DOI : https://doi.org/10.1038/s41467-021-24952-6

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

This article is cited by

An all-in-one seismic phase picking, location, and association network for multi-task multi-station earthquake monitoring.

Communications Earth & Environment (2024)

Deep learning forecasting of large induced earthquakes via precursory signals

• Vincenzo Convertito
• Fabio Giampaolo
• Francesco Piccialli

Scientific Reports (2024)

Cascade and pre-slip models oversimplify the complexity of earthquake preparation in nature

• Patricia Martínez-Garzón

The Role of Machine Learning in Earthquake Seismology: A Review

• Anup Chitkeshwar

Archives of Computational Methods in Engineering (2024)

Machine learning-based multipath modeling in spatial domain applied to GNSS short baseline processing

• Yuanxin Pan
• Gregor Möller
• Benedikt Soja

GPS Solutions (2024)

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

• Frontier Letter
• Open access
• Published: 28 February 2024

Recent advances in earthquake seismology using machine learning

• Hisahiko Kubo   ORCID: orcid.org/0000-0002-1933-7899 1 ,
• Makoto Naoi   ORCID: orcid.org/0000-0001-9488-9266 2 &
• Masayuki Kano   ORCID: orcid.org/0000-0002-7288-4760 3  

Earth, Planets and Space volume  76 , Article number:  36 ( 2024 ) Cite this article

2383 Accesses

12 Altmetric

Metrics details

Given the recent developments in machine-learning technology, its application has rapidly progressed in various fields of earthquake seismology, achieving great success. Here, we review the recent advances, focusing on catalog development, seismicity analysis, ground-motion prediction, and crustal deformation analysis. First, we explore studies on the development of earthquake catalogs, including their elemental processes such as event detection/classification, arrival time picking, similar waveform searching, focal mechanism analysis, and paleoseismic record analysis. We then introduce studies related to earthquake risk evaluation and seismicity analysis. Additionally, we review studies on ground-motion prediction, which are categorized into four groups depending on whether the output is ground-motion intensity or ground-motion time series and the input is features (individual measurable properties) or time series. We discuss the effect of imbalanced ground-motion data on machine-learning models and the approaches taken to address the problem. Finally, we summarize the analysis of geodetic data related to crustal deformation, focusing on clustering analysis and detection of geodetic signals caused by seismic/aseismic phenomena.

Graphical Abstract

Introduction

Recently, machine learning (ML) technology, including deep learning (DL), has made remarkable progress in various scientific fields, including earthquake seismology, producing vast research findings. Here, we review the applications of ML in several fields of earthquake seismology and discuss the strengths and difficulties of using ML. We focus on ML applications in earthquake catalog development, seismicity analysis, ground-motion prediction, and geodetic data related to crustal deformation. Figure  1 summarizes the topics reviewed in this paper. Notably, our focus is on exploring the applications of ML in individual fields rather than providing a comprehensive review of ML applications in the entire field of study. Excellent review articles have been provided for the reader’s reference on the recent applications of ML in solid earth geosciences by Bergen et al. ( 2019 ), seismology by Kong et al. ( 2019 ) and Mousavi and Beroza ( 2022 , 2023 ), microseismic monitoring with small signals by Li ( 2021 ) and Anikiev et al. ( 2023 ), and analysis of Global Navigation Satellite System (GNSS) data by Siemuri et al. ( 2022 ).

Topics reviewed in this study

Application in earthquake catalog development

The technique of detecting seismic waveforms, determining hypocenters, and cataloging them from seismogram records is an area driving the use of ML in seismology. This section reviews the recent developments in ML applications related to earthquake cataloging, focusing on improving individual tasks in its pipeline, similar waveform searching, focal mechanism analysis, and paleoseismic record analysis.

Importance of improving earthquake cataloging methods and contribution of machine learning techniques

Considering that the earthquake size distribution follows a power law, improving event detectability drastically increases the number of analyzable events, enhancing the spatiotemporal resolution in seismicity analysis. Some studies have emphasized the role of small events in the earthquake-generation process, highlighting the importance of making complete earthquake catalogs. For example, Mignan ( 2014 ) conducted a meta-analysis of 37 foreshock studies and found that interpretation of foreshock activity depends on the completeness magnitude of the used earthquake catalogs. Trugman and Ross ( 2019 ) showed that precursory seismicity is more ubiquitous than previously understood in southern California using a highly complete catalog. Naoi et al. ( 2015 ) reported that on-fault seismicity comprises only very small (M < – 2) events in a gold mine. The technology to produce high-quality seismic catalogs from seismic waveforms contributes to achieving a smaller completeness magnitude, as well as improving station density and seismometer sensitivity.

To detect small seismic events with waveforms buried in noise, the template matching technique (Gibbons and Ringdal 2006 ; Shelly et al. 2007 ; Peng and Zhao 2009 ), whereby additional events with waveforms similar to template waveforms are explored, has frequently been used. This technique often reduces the completeness magnitude by approximately 1, resulting in an increase in the number of detected events by an order of magnitude (Ross et al. 2019a ). Although this approach works well, its application is usually limited to short-period data (e.g., preceding large earthquakes) owing to the high computing costs. Parallel computation using many Graphics Processing Units (GPUs) enables the application of this method to a larger dataset (Ross et al. 2019a ); however, such an environment is not readily available, resulting in the rarity of such cases.

The application of ML techniques to efficiently develop high-quality earthquake catalogs has rapidly progressed since approximately 2018. The application of DL to event detection and arrival time reading problems is driving this study area, and its performance has continuously improved by adopting new architectures and modules inspired by rapid advancements in the ML field. ML provides more accurate and noise-resistant auto-processing methods than conventional automatic processing, resulting in highly complete earthquake catalogs. Similar to template matching, the recent ML-based method often increases the number of events in catalogs by tenfold compared with conventional approaches (Tan et al. 2021 ; Ma and Chen 2022 ) with a much lower computation cost (Perol et al. 2018 ). By benefiting from the developments in this field, many studies exploring seismicity with high resolution have been published (Liu et al. 2020 ; Park et al. 2020 ; Ross et al. 2020 ; Wang et al. 2020a ; Baker et al. 2021 ; Li et al. 2021 ; Jiang et al. 2022 ; Gong et al. 2022a ; Chen et al. 2022a ). This improvement in catalog quality is expected to improve earthquake forecasts based on seismicity models (Mancini et al. 2022 ).

The models trained on large data sets in some of these studies are publicly available (Ross et al. 2018b ; Zhu and Beroza 2019 ; Mousavi et al. 2020 ), making it easy to benefit from this development. These models are versatile (Cianetti et al. 2021 ) and can be applied to many datasets. Even when the scale of the target earthquakes or the observation frequency bands differs significantly from the training data, transfer learning and fine-tuning (Pan and Yang 2010 ; Bozinovski 2020 ), in which an ML model trained on one task is recalibrated for a related task, helps analyze data (Titos et al. 2020 ; Chai et al. 2020 ; Lara et al. 2021 ; Lapins et al. 2021 ; Kim et al. 2023 ), likely due to the self-similarity of earthquakes. These techniques also enable the application of ML to project data where a large amount of training data is difficult to prepare. The application of ML has also been reported in cataloging development problems related to acoustic emissions (AEs) in laboratories (Trugman et al. 2020 ; Li et al. 2022c ). Many studies have published ML-based packages that handle the entire catalog development procedure throughout the event detection from continuous waveform records, arrival time picking, phase associations, and hypocenter locations (Walter et al. 2021 ; Zhu et al. 2022c , a ; Retailleau et al. 2022 ; Shi et al. 2022 ; Zhang et al. 2022c ; Dokht et al. 2022 ). An environment to use ML for seismicity analysis is being developed as a basic technology in seismology.

Improvement of individual tasks in cataloging pipelines

Developing an earthquake catalog from seismic waveform data is typically divided into the following steps: (1) event detection from continuous waveform record; (2) arrival time reading; (3) phase association; and (4) hypocenter determination (Fig.  2 ). This is simple modeling, and each process can be skipped or integrated. For example, event detection is often skipped, and phase picking is performed directly on continuous data. Moreover, to improve performance, a denoising phase (Tibi et al. 2022 ) can be added before the detection process, and a quality control phase (Tamaribuchi et al. 2021 ) can be added after the hypocenter determination.

Typical pipeline for earthquake catalog development

Application of ML techniques has been attempted in most of these processes: their performance significantly outperforms conventional methods. Several models trained on regional or global datasets are publicly available, especially for event detection and arrival time picking problems. These performances have been systematically compared (Cianetti et al. 2021 ; Münchmeyer et al. 2022 ; García et al. 2022 ), and some studies opened datasets for benchmarking the model performance (Mousavi et al. 2019a ; Magrini et al. 2020 ; Michelini et al. 2021 ; Woollam et al. 2022 ). Mai and Audet ( 2022 ) developed software to generate labeled datasets, while Woollam et al. ( 2022 ) developed a common application programming interface to access various ML models in those tasks.

Event detection/classification and arrival time picking

Certain algorithms, including the ratio between short-term average (STA) and long-term average (LTA) (Allen 1978 ), have been conventionally used to extract transient signals from a continuous seismic record. The ML technique solves this problem with higher performance, and the application can be expanded to extract many types of signals.

The event detection task can be treated as a problem of classifying cutout seismic waveforms into two (typically earthquakes and noises) or more categories. This issue has been addressed by various ML algorithms, including support vector machine, decision tree, logistic regression, and neural network, including DL (Reynen and Audet 2017 ; Meier et al. 2019 ; Tang et al. 2020 ; Albert and Linville 2020 ; Kim et al. 2021b ; Chakraborty et al. 2022 ; Murti et al. 2022 ). The target of such classifiers can be expanded to various phenomena by preparing appropriate training data. For example, it can be used to classify/identify surface waves (Chai et al. 2022 ), volcanic earthquakes (Bueno Rodriguez et al. 2022 ; Canário et al. 2020 ; Lara et al. 2021 ), moonquakes (Civilini et al. 2021 ), low-frequency earthquakes (Nakano et al. 2019 ; Rouet-Leduc et al. 2020 ; Thomas et al. 2021 ; Takahashi et al. 2021 ; Chen et al. 2023 ), and mining-induced earthquakes/bastings (Linville et al. 2019 ; Tibi et al. 2019 ; Peng et al. 2021 ; Wang et al. 2023b ). Applications in distributed acoustic sensing, which require efficient processing of large amounts of data, have also been reported (Hernandez et al. 2022 ). Researchers have used simple Convolutional Neural Network (CNN) models (Zhang et al. 2020a ; Majstorović et al. 2021 ) and the latest architecture/modules, such as Recurrent Neural Network (RNN) (Mousavi et al. 2019c ), attention (Hou et al. 2023 ), transformer (Mousavi et al. 2020 ), multi-feature fusion networks (Kim et al. 2021a ), graph-partitioning based CNN (Yano et al. 2021 ), Capsule Neural Network (Saad and Chen 2022 ), Residual Neural Network (ResNet) (Li et al. 2022b ) and inceptions (Jia et al. 2022 ), contributing to improved performance. Moreover, some studies have aimed to enhance the interpretability of DL, often treated as black boxes (Kong et al. 2022 ; Majstorović et al. 2022 ).

The form of the input/output data of such classifiers can be designed flexibly. For example, one may design a classifier to output a scalar value corresponding to the classification result for a cutout waveform (Zhang et al. 2020a ). Applying this approach to moving windows (Ross et al. 2018b ; Dokht et al. 2019 ; Zhu et al. 2019a ) enables the generation of a time series for the probability of a seismic signal. An architecture that outputs vectors with the same number of samples as input waveforms (Mousavi et al. 2020 ) allows for determining each data point’s probability. This likely offers advantages when target events have widely varying durations or when many events occur frequently in a short period. Given that DL techniques have been developed for image processing, many studies have used two-dimensional data with time–frequency representation as input data (Dokht et al. 2019 ; Mousavi et al. 2019c ; Lara et al. 2021 ; Saad et al. 2021 ). Nevertheless, one-dimensional CNNs have also been used frequently owing to their simple pre-processing of waveform data and low computational cost (Nakano and Sugiyama 2022 ). In addition to the simplest case using a single-channel waveform at a single station, multi-channel or multi-station waveforms can easily be inputted into a model. When using a single-channel waveform, a large amount of training data can be prepared more easily, and a model independent from specific observation settings can be obtained, resulting in highly versatile models. When using multi-channel or multi-station data as input, models will likely achieve high detectability and robustness for local noises (Yang et al. 2021a ; Yano et al. 2021 ). Another way to use multi-station data for event detection is to simultaneously estimate hypocenter locations with event detection by creating a classifier based on coherence images of characteristic functions estimated from waveforms (Mosher and Audet 2020 ). This approach is closely related to the topics discussed in the “ Earthquake location ” subsection.

Preparing training data is one of the most important tasks in developing a classifier using supervised learning. For example, when performing binary classification of earthquake and noise, it is necessary to prepare training data for both classes. Compared with preparing seismic event datasets, for which a vast amount of analysis results (including manual processing) has been accumulated, preparing noise waveform datasets with a wide variety of properties may be more difficult. A potential approach to solve this problem is waveform classification using unsupervised learning, which can be generalized as a clustering problem for seismometer records. Early studies in this field attempted clustering based on human-selected features with sufficiently reduced dimensions before inputting into ML models (Chamarczuk et al. 2020 ; Johnson et al. 2020 ). Recently, more effective clustering has been attempted by automatically extracting features directly from unprocessed datasets, such as seismic waveforms or their spectrograms (Seydoux et al. 2020 ; Yin et al. 2022b ). This approach is used for various purposes, including the detection/classification of earthquakes (Seydoux et al. 2020 ; Steinmann et al. 2022a ), quarry blasts and rockfalls (Hammer et al. 2013 ), volcanic earthquakes/tremors (Esposito et al. 2008 ; Unglert and Jellinek 2017 ; Soubestre et al. 2018 ; Ren et al. 2020 ; Ida et al. 2022 ), surface freezing and thawing (Steinmann et al. 2022b ), reflection wave (Ali et al. 2022 ), and seismic waves radiated in glaciological processes (Jenkins et al. 2021 ).

Incidentally, in the initial process of earthquake catalog development, the analyst typically identifies a window involving a seismic waveform and then measures arrival time in a cutout waveform. However, this event detection phase can be skipped, resulting in the arrival time measurements being performed directly on the entire continuous waveforms (Park et al. 2020 ; Heck et al. 2022 ; Retailleau et al. 2023 ). The fraction of false detections of arrival time measurements and the efficiency of the subsequent phase association calculations likely determine which of the two approaches is effective.

Automatic arrival-time reading, for which the method using the autoregression model and Akaike Information Criteria (Takanami and Kitagawa 1988 ) is famous, could not achieve the accuracy of manual processing until the advent of methods based on supervised DL, i.e., neural phase pickers (Park et al. 2023 ), especially for S-wave arrival times. Although the specific problem of “prediction inconsistency” has been reported (Park et al. 2023 ), DL-based pickers outperform traditional algorithms and achieve picking accuracies similar to those of skilled analysts (Mousavi and Beroza 2023 ), and the published models have been used in many studies. Similar to event detection/classification problems, state-of-the-art modules and architectures, such as RNN (Zhou et al. 2019 ), attention (Liao et al. 2021 , 2022a ; Li et al. 2022a ), transformer (Mousavi et al. 2020 ), and edge convolutional module (Feng et al. 2022b ), were continuously incorporated into the models to improve their performance. In this approach, a model takes seismic waveforms as inputs and outputs the arrival times as scalar values (Ross et al. 2018a ) or a time series of probability values with a peak at a picked arrival time (Zhu and Beroza 2019 ; Mousavi et al. 2020 ). Xiao et al. ( 2021 ) showed that the performance of such a model for low SN waveforms can be improved by learning the waveform similarities. Although most previous studies on arrival time reading problems were conducted on body waves, applications have begun to read subsequent waves, such as reflected waves, to use structure imaging (Ding et al. 2022 ; Kato 2023 ).

Similar to the event detection problem, arrival time reading models based on DL allow a flexible input/output data format. While some studies have used single-station waveforms (Ross et al. 2018a ; Zhu and Beroza 2019 ; Woollam et al. 2019 ; Wang et al. 2019a ; Zheng et al. 2020 ; Mousavi et al. 2020 ; Liao et al. 2021 ; Tokuda and Nagao 2023 ), others have applied multiple station records (Zhu et al. 2022c ; Li et al. 2022c ; Chen and Li 2022 ; Feng et al. 2022b ; Sun et al. 2023 ). As both event detection and phase picking use seismic waveform as input data, they can be processed simultaneously in a single model using multitask learning, which has multiple outputs, including event and arrival time probabilities (Liao et al. 2022a ; Mousavi et al. 2020 ).

A common challenge in event detection and arrival time reading based on supervised learning is the insufficient training data for large events, which naturally results from the power-law size distribution of earthquakes. When training data are generated from past observation data, attention should be paid to the performance for unknown events, which have locations, sizes, and focal mechanisms not occurring previously. Data augmentation can complement this problem, and effective augmentation methods were investigated by considering the characteristics of seismic data (Zhu et al. 2020 ). Zhang et al. ( 2021b ) trained a phase-picking model on an augmented dataset based on only ten hydraulically induced events, demonstrating that the resultant model can create event catalogs from seismic records in different stages and projects. Additionally, Generative Adversarial Networks (GANs; Goodfellow et al. 2014 ) were attempted to generate training data (Li et al. 2020b ; Wang et al. 2021 ) and perform feature extraction (Li et al. 2018 ). Marano et al. ( 2023 ) reviewed the use of GANs in seismology, including data augmentation.

Phase association

When all arrival time candidates in a specific time range are used to determine the hypocenters, the location accuracy often reduces significantly due to misidentification of the arrival time (false positives). This problem can be prevented in the location process by removing candidates with large residuals in iterations (Waldhauser and Ellsworth 2000 ; Naoi et al. 2018 ). For efficient analysis, some studies exclude false positives by introducing an additional process called “phase association”, during which arrival time candidates are associated with a single earthquake. If the association is computationally efficient and accurate, the existence of many false positives is not problematic in the subsequent processing. This allows analysts to prefer “recall” to “precision” in the arrival time reading process, possibly resulting in a better catalog with a smaller number of missed events. Therefore, the performance of the phase association process is important in the earthquake cataloging process.

A popular association method that works well is calculating the theoretical arrival times for a given hypocenter and searching for the hypocenter with many compatible arrival time candidates using a grid search (Draelos et al. 2015 ). Recently, the REAL package (Zhang et al. 2019 ) efficiently processed this task and has, thus, been frequently used. ML-based methods have also been developed to address this issue. For example, Yu and Wang ( 2022 ) used a neural network to solve the grid search problem efficiently. The association methods based on RNN (Ross et al. 2019b ), Graph Neural Networks (McBrearty and Beroza 2023 ), Bayesian Gaussian mixture models (Zhu et al. 2022b ), and Physics-Informed Neural Networks (PINNs; Ross et al. 2023 ) have also been proposed. Methods based on binary classification of waveform pairs (McBrearty et al. 2019 ) and deep metric learning (Dickey et al. 2020 ) have been attempted to directly determine from seismic waveforms whether they are generated by the same earthquake.

Without the phase association process, the hypocenters can be directly determined by migrating/scanning the time series of arrival time probabilities, although the computational cost is high. This topic is discussed in the next subsection.

Earthquake location

After the arrival times of many stations are obtained for an earthquake, the hypocenter of the event can be inversely solved with high accuracy and small computing cost, as is generally performed in seismology (Geiger 1912 ). Some studies have to address this “travel time-based” location problem using ML techniques, including random forest (Saad et al. 2022 ; Chen et al. 2022c ) and fully connected neural networks (Anikiev et al. 2022 ). These approaches are possibly effective for earthquake early warning (EEW), where hypocenters must be roughly estimated using a small number of arrival time candidates as early as possible (Saad et al. 2022 ).

Another approach to hypocenter determination is the “waveform-based” location method, which does not require phase picking and identifies the hypocenter based on migration and stacking of waveforms or characteristic functions calculated from waveforms at many stations. Although this approach involves high computation costs, it contributes to detecting small earthquakes with waveforms buried in noise (Li et al. 2020a ). Compared with characteristic functions, such as STA/LTA (Grigoli et al. 2013 ) or kurtosis (Langet et al. 2014 ) time series used in previous studies, the probability trace estimated using a DL phase picker (Liao et al. 2022b ) likely significantly reduces the false positive rate. This approach is often adopted in open packages that handle the entire cataloging process from continuous seismic records to hypocenter locations (Zhu et al. 2022c ; Shi et al. 2022 ). Zhu et al. ( 2022c ) incorporated this approach into a single DL model that takes the seismic waveform as input and outputs a hypocenter (end-to-end processing) to minimize discarded information during the cataloging procedure. Other studies have applied ML techniques to determine hypocenters from coherence/migration images. Zhang et al. ( 2022d ) proposed a DL-based approach that used diffraction stacking images of seismic waveforms and solved a regression problem to determine the hypocenter from these images. Wu et al. ( 2019 ) proposed a method to determine hypocenters using deep reinforcement learning by reconstructing a three-dimensional energy field constructed from observed waveforms.

As described above, in the hypocenter determination problem, the optimal hypocenter is generally searched using arrival times or related information (e.g., characteristic function) extracted from the waveforms at multiple stations. However, the later phases of the seismic waveform contain additional information on Green’s function between the hypocenter and station. If such information could be used effectively, hypocenters can be determined based even on single-station data. As a simple and conventional approach, a hypocenter can be roughly determined by estimating the direction of the ray path that can be estimated from the P-wave-particle motion and hypocentral distance that can be estimated from the arrival time difference between different phases (e.g., P and S phases) (Frohlich and Pulliam 1999 ). ML approaches are expected to improve the performance using information discarded in the conventional method. Gutierrez et al. ( 2019 ) estimated the azimuth direction of incident seismic waves using support vector machines. Mousavi and Beroza ( 2020 ) and Ristea and Radoi ( 2022 ) estimated hypocenter locations or related parameters using DL from a single station waveform. Perol et al. ( 2018 ) proposed ConvNetQuake, which roughly determines hypocenters from a single station waveform by solving the classification task for the divided analysis regions using DL; this method has been improved in subsequent studies (Lomax et al. 2019 ; Tous et al. 2020 ; Bai and Tahmasebi 2021 ).

As a natural consequence of the above discussion, the ideal way to determine the hypocenter is to use as much information as possible from the full seismic waveforms at multiple stations. To achieve this, Kriegerowski et al. ( 2019 ), Zhang et al. ( 2020b ), Shen and Shen ( 2021 ), and Münchmeyer et al. ( 2021 ) developed DL models to determine source parameters from seismic waveforms at multiple channels using the waveforms of the past events as training data. The introduction of a graph-based approach improved the performance of these tasks (van den Ende and Ampuero 2020 ; Zhang et al. 2022e ). Piras et al. ( 2022 ) used a Bayesian approach to estimate a hypocenter from the observed spectra of seismic waveforms. They used a neural network to map spectral-based features from the hypocenter for the posterior distribution sampling; this is prohibitively expensive using conventional methods, especially under a three-dimensional heterogeneous density–velocity model. Although these methods work well, observation data are generally incomplete, limiting the effectiveness for events with hypocenter locations and focal mechanisms not included in the training data. To overcome this challenge, some studies have attempted training based on synthetic waveforms (Wang et al. 2022c ; Sugiyama et al. 2021 ; Wamriew et al. 2022 ; Vinard et al. 2022 ), including data generated with semi-supervised GAN (Feng et al. 2022a ).

To address the hypocenter determination problem, attempts have been made to use PINNs, which incorporate theoretical constraints, including physical laws in loss functions (Smith et al. 2021 ; Izzatullah et al. 2022 ). This additional constraint is advantageous when analyzing unknown data (not included in the training data). PINNs are trained to solve supervised learning tasks while satisfying any physical law described by partial differential equations (PDEs) by embedding the PDEs into the loss function of a neural network using automatic differentiation (Raissi et al. 2019 ; Karniadakis et al. 2021 ). PINNs can represent continuous solutions of PDE without discretization and can be applied to various types of forward and inverse problems (e.g., seismic tomography) (Chen et al. 2022b ; Agata et al. 2023 ); consequently, they have received significant attention in science, including geophysics.

Quality control

For the hypocenters obtained using the pipelines described above, the solution quality is generally controlled by thresholds for some parameters, including the arrival time residuals in the inversion, to ensure that the false positive rate is sufficiently small. Ideally, this strategy can separate correct and incorrect solutions; however, it usually results in many false positives and false negatives. The tradeoff between the false positive rate and the number of solutions is possibly an obstacle to developing a high-quality catalog. Nonlinear classification using ML can potentially solve this problem. For example, Tamaribuchi et al. ( 2021 , 2023 ) used ensemble learning to exclude false positives from hypocenters obtained by automatic processing routinely operated by the Japan Meteorological Agency. Tamaribuchi et al. ( 2023 ) also used the CNN model by Ross et al. ( 2018b ) to screen automatic phase-picking results. Beaucé et al. ( 2019 ) and Herrmann et al. ( 2019 ) used ML to remove false positives from energy-based seismic detection results and event candidates detected by matched filter analysis, respectively.

Similar waveform searching

As mentioned previously, cross-correlation-based techniques such as template matching and autocorrelation methods, which explore similar waveforms in continuous seismic records, increase the number of available events (Gibbons and Ringdal 2006 ; Shelly et al. 2007 ; Peng and Zhao 2009 ; Kato et al. 2012 ; Chamberlain et al. 2018 ). These methods effectively extract small signals buried in noise and are vital in the analysis of seismic activity. However, these methods require cross-correlation calculation among all combinations of the analyzed event waveforms, resulting in high computational costs. For example, the computational cost of autocorrelation analysis, in which brute-force calculations are performed on continuous waveform records, is O(N 2 ), limiting the analyzable period. Even for the template-matching problem, large-scale dataset requires the parallel computation of many GPUs (Ross et al. 2019a ). Furthermore, template matching using synthetic or empirical synthetic waveforms (Chamberlain and Townend 2018 ; Ide 2021 ) potentially contributes to better catalog development. Hence, an efficient similar waveform search is an important technique in catalog development.

The ML technique potentially reduces the computational cost in similar waveform searches. For example, Skoumal et al. ( 2016 ) proposed the Repeating Signal Detector method, which detects seismic signals through unsupervised learning and enables template matching without template preparation before applying the method or brute force calculation. Ganter et al. ( 2018 ) proposed a similar waveform search based on ML to generate templates satisfying the alternate null hypothesis, which incorporates the possibility of false positives. Kumar et al. ( 2022 ) suggested that using the dynamic time warping technique instead of cross-correlations possibly improves the performance of similar waveform searching, including template matching analysis and unsupervised waveform clustering.

As an efficient similarity search technique, approximate nearest neighbor searches have been used in the field of image and speech recognition and have several applications in seismology (Yoon et al. 2015 ; Tibi et al. 2017 ). For example, Yoon et al. ( 2015 ) applied the Locally Sensitive Hashing (LSH) technique based on random substitution to seismic waveforms and developed a method called FAST that can efficiently search for similar waveforms by hashing the binary fingerprint generated from the waveform spectrogram. Using this method, Yoon et al. ( 2015 ) performed a brute-force calculation for a single channel, six months of waveform records. Scotto di Uccio et al. ( 2022 ) reported that a similar waveform search using FAST potentially complements events not detected using DL-based methods, although they achieved the best performance using a template-matching technique based on a template catalog developed using a DL-based method. The FAST algorithm has been used for induced earthquakes due to hydraulic fracturing and wastewater injection (Yoon et al. 2017 ), induced earthquakes in gas fields (Scala et al. 2022 ), swarms (Festa et al. 2021 ), foreshocks (Yoon et al. 2019b ), and volcano seismicity (Garza-Girón et al. 2023 ).

The LSH-based method is far more scalable than waveform correlation methods. Additionally, the FAST performance has been enhanced by avoiding false positives using multiple channel/station results (Bergen and Beroza 2018 ), improving fingerprint (Bergen et al. 2016 ; Bergen and Beroza 2019 ) and hash table treatment through careful tunings, such as partitioning (Rong et al. 2018 ). Consequently, Yoon et al. ( 2019a ) successfully applied this method for 11 stations, 27 channels, and 6–11 years of continuously recorded seismic data. However, the computational cost to process large-scale data remains high, limiting the size of the analyzable dataset. The DL-based method, which is beginning to be used in the field of image and audio recognition, is a potential solution to reduce the computational cost. Naoi and Hirano ( 2024 ) developed a method to convert a seismic waveform into a 64-bit binary code and applied it to 16-channel, 30-min 10 MS/s continuous AE records (the number of data points in each channel corresponds to 5.8-year continuous waveform records of 100 Hz sampling typically adopted in seismic observation) obtained from laboratory experiments. They showed that the hashing-based template matching and autocorrelation problems of these datasets can be solved in a realistic computation time.

In summary, besides improving processing pipelines in earthquake cataloging, ML enables efficient similar waveform searches that provide additional information, making it feasible to apply these methods to large-scale data, for which the computational cost has been an obstacle.

Focal mechanism analysis

Focal mechanism analysis requires polarities of P-wave first motions, and moment tensor analysis requires amplitude information or waveform inversion. The necessity of manual processing and the insufficient accuracy of techniques that automatically obtain these data have impeded the analysis of many small events.

The task of polarity reading for the P-wave first motions can most simply be treated as a binary classification problem (Ross et al. 2018a ) that can be readily solved with DL. Similar to the arrival time reading problem, this approach exhibits high performance and markedly increases available focal mechanism solutions. For example, Uchide ( 2020 ), Cheng et al. ( 2023 ), and Tanaka et al. ( 2021 ) used DL to read first motion polarities of seismic waveforms or AE data in Japan, southern California, and laboratory hydraulic fracturing experiments, respectively, obtaining tens to hundreds of thousands of focal mechanisms. Naoi et al. ( 2022 ) automatically read the timing of the initial maximum amplitude using DL; the obtained amplitudes were used to solve moment tensors for laboratory AEs. In these analyses, waveforms from a single station were typically used as input data to a DL network; additionally, multi-station approaches have been explored (Tian et al. 2020 ). Although it is not as active as the problem of arrival time reading, the generalizability of these models has also been studied (Hara et al. 2019 ). Furthermore, unsupervised clustering using DL can identify waveforms with different initial polarities without training data (Mousavi et al. 2019b ). The increased number of available focal mechanisms improves spatiotemporal resolution for mechanism-based analyses, such as stress tensor inversion (Uchide et al. 2022 ; Qin et al. 2022b ). ML is also helpful in interpreting such a large number of solutions. Kubo et al. ( 2023 ) applied nonlinear graph-based dimensionality reduction to a large-scale dataset of moment tensor solutions, objectively obtaining a comprehensive image of the spatiotemporal characteristics of seismicity.

P-wave first motion polarities should show the same pattern for events with similar hypocenters and focal mechanisms. Using this property, Skoumal et al. ( 2023b ) improved the number and accuracy of focal mechanism solutions employing ML to complement P-wave first motion data at stations not used in the mechanism analysis due to the insufficient signal-to-noise ratio. The polarity of such stations can also be complemented using the waveform correlation technique (Shelly et al. 2016 ). By combining these methods, Skoumal et al. ( 2023a ) successfully estimated many focal mechanisms for events in the southeastern San Francisco Bay Area.

Regarding ML applications in focal mechanism analysis, in addition to improving the polarity reading and amplitude estimation, some studies have attempted to rapidly estimate the focal mechanisms for EEW (Kuang et al. 2021 ; Steinberg et al. 2021 ; Monterrubio-Velasco et al. 2022 ). Physics-Guided Neural Networks, which use ML for inverse analysis of the constrained solution space of the focal mechanisms, have been investigated (Zhang et al. 2021a ). Future applications are expected to include other physics-constrained neural networks, such as PINNs and Physics-Encoded Neural Networks (Faroughi et al. 2022 ).

Paleoseismic record analysis

ML-based techniques for processing digitized waveform data have been rapidly developed. However, such digital records are available only for the last quarter of the period since the beginning of seismometer observations (Wang et al. 2019b ). Past seismic activity should be investigated using analog records to assess the history of long-period seismic cycles. Although there are few examples, some important attempts have been reported. Furumura et al. ( 2023 ) applied a CNN network to digitize analog, strong‐motion seismographs recorded on smoked paper. Wang et al. ( 2019b , 2022d ) performed event detection, arrival time reading, magnitude estimation, and hypocenter determination on continuous Develocorder film images obtained from the Rangely seismic control experiment by the United States Geological Survey, generating a seismic event catalog. Kaneko et al. ( 2021 , 2023 ) also identified past low-frequency earthquake records among analog records using DL.

Current achievements and future prospects

As described in this section, ML is utilized in nearly every step of the earthquake catalog development pipeline. In particular, conventional automatic processing methods have not reached the accuracy of manual processing for regarding event detection, arrival time reading, and P-wave first-motion polarity reading. This has impeded the efficient processing of large amounts of data. DL models have demonstrated outstanding performance and have greatly improved the accuracy of automatic processing. Although the limited training data for large earthquakes and events in the regions or with focal mechanism types that have not previously occurred are likely to pose inherent challenges in these tasks, the outstanding performance of the DL approach has led to progress in related research. Some trained models are publicly available and have been applied to a wide array of data, either as-is or after optimization through transfer learning or retraining. The required computational cost is sufficiently small in the prediction process, facilitating the easy use of these models and greatly advancing seismicity analysis. Their performance will likely continue to improve via architecture modification and model development considering consistency across multiple station records.

Regarding the phase association and location, the adoption of ML technology is in its initial stages. The benefits and need for ML in these areas may remain unclear and warrant further research. Although applications to quality control problems also remain in the early stages of exploration, the ML technique is likely to be highly suitable for this task. DL has also been utilized for similar content search problems in other research fields and is expected to contribute to efficient similar waveform searching in seismology. The application to paleoseismic records is also in its infancy; however, ML technology, which has achieved great success in the image recognition problem, is likely to be well suited to address the related tasks efficiently. Effectively using ML in this domain could boost efforts to preserve paper records.

As described above, ML improves (or is expected to improve) the performance of individual tasks of the catalog development pipeline. However, in catalog development, it is not only necessary to enhance individual processes but also to optimize the overall performance. The catalog development pipeline, which includes similar waveform search processes, offers numerous options, making its optimization complex. Optimizations using an end-to-end approach, as Zhu et al. ( 2022c ) suggested, are a potential approach to address this issue.

The ML model demonstrates generalizability and provides good results when analyzing seismic records from regions that differ from the training data. However, for widely used public datasets, sharing retrained models with more relevant data is beneficial. Preparing/sharing such models as a research infrastructure will enhance the related study. Datasets for retraining should ideally be publicly accessible and easily applicable for future model development. For example, the Seisbench toolbox (Woollam et al. 2022 ) provides a dedicated software environment for this purpose. To encourage such efforts, a clear policy on the secondary distribution of public data is essential.

Application to earthquake risk evaluation and seismicity analysis

In addition to contributing to earthquake catalog development, ML has been applied to seismicity analysis, including risk evaluation. Here, we summarize the related topics, including the applications in laboratory and induced earthquakes. Notably, we have not provided a comprehensive review of studies related to earthquake forecasting using ML (e.g., Ridzwan and Yusoff 2023 ). Mignan and Broccardo ( 2020 ) provided an excellent review of the current achievements in earthquake prediction using artificial neural networks.

Rupture forecasting of laboratory earthquakes

Frictional experiments in the laboratory are relatively successful targets in predicting the occurrence of macroscopic failures. In fracture or slip experiments in a laboratory, significant precursors are often observed, such as increased AE activity, decreased b-values, and changes in fractal dimensions (e.g., Lei and Satoh 2007 ; McLaskey and Kilgore 2013 ).

In slip experiments, it is possible to generate many seismic cycles, providing rich training data for ML to predict the timing and magnitude of macroscopic failures. This topic was triggered by Rouet-Leduc et al. ( 2017 ), who analyzed continuous AE records obtained during double-direct shear experiments. Thereafter, the passive and active AE records obtained in this type of experiment have been actively studied. In a pioneering study, Rouet-Leduc et al. ( 2017 ) developed a random forest model that could predict the timing of macroscopic slip from AE waveform data continuously recorded during the experiment. They showed that the time to macroscopic slip could be predicted well not only from the data immediately before the macroscopic failure, but also throughout the entire period of seismic cycles, including their early stage. Their prediction was based only on window data at a particular time without data from the entire cycle. This suggests that the AE waveforms contain information on the preparation process for future macroscopic failure throughout the seismic cycle. Rouet-Leduc et al. ( 2018b ) applied a similar method to the field data and successfully reproduced surface displacements from continuous seismic data in the Cascadia subduction zone, indicating that tremors in this region occurred nearly at all times.

The issue raised by Rouet-Leduc et al. ( 2017 ) was treated in a Kaggle competition to improve prediction algorithms (Johnson et al. 2021 ), resulting in several published methods (Brykov et al. 2020 ; Laurenti et al. 2022 ). Related to this problem, Bolton et al. ( 2019 ) examined whether the unsupervised clustering technique could identify precursors for similar experimental data, while Jasperson et al. ( 2021 ) further improved the prediction accuracy by combining unsupervised clustering and attention networks. Rouet-Leduc et al. ( 2018a ) solved a similar issue using a gradient-boosted tree and showed that the shear stress on the fault plane (i.e., the frictional state) could be reproduced from continuous AE recordings. They found that the variance in the waveform amplitude, a function of the average energy per unit of time, is the parameter that primarily contributes to the prediction. Hulbert et al. ( 2019 ) also found that the event duration, that is, the slip mode (slow or fast event), could be predicted in the experimental system where slow events possibly occur. Bolton et al. ( 2020 ) showed that fault slip velocity and frictional contact area per unit fault volume govern the variance of AE waveform amplitudes.

Lubbers et al. ( 2018 ) reported that similar predictions are possible using AE catalogs comprising magnitude and time without using continuously recorded waveform data. Their prediction performance improved when using a more complete catalog (i.e., a catalog with a low completeness magnitude), possibly corresponding to the fact that the completeness of the earthquake catalog is important in detecting foreshock activities preceding large earthquakes (Mignan 2014 ; Trugman and Ross 2019 ). Shokouhi et al. ( 2021 ) and Shreedharan et al. ( 2021 ) confirmed that temporal changes in shear stress and time to macroscopic failure could be predicted using an active source, that is, transmission test data. This is not surprising given that many previous studies have indicated that fault transmission waves can be used to monitor fault conditions (Kendall and Tabor 1971 ; Nagata et al. 2008 ). To address a similar prediction problem based on active source data, Borate et al. ( 2023 ) successfully improved model performance using a PINN model incorporating physical constraints related to transmission waves, particularly when the training data size was small.

Inspired by the double direct shear experiment, ML predictions have also been attempted in an analog experiment simulating a subduction zone (Rosenau et al. 2017 ). In these studies, issues of predicting the time to macroscopic slip (Corbi et al. 2019 ), determining whether slip will occur immediately after the analyzed data window (Corbi et al. 2020 ), and predicting future surface velocity fields (Mastella et al. 2022 ), were addressed based on surface deformation data, by assuming the future application to the actual GNSS data.

As described above, ML models trained on data during many cycles of laboratory experiments have successfully predicted macroscopic ruptures and estimated the conditions of fault planes and medium. The application to double direct shear experiments data, pioneered by Rouet-Leduc et al. ( 2017 ), has driven this field. Future efforts to apply a similar approach to various laboratory experiments might identify important signals previously overlooked.

In laboratory data analysis, ML techniques reveal the physics of fault deformation in various ways besides the prediction issue. For example, Trugman et al. ( 2020 ) used ML to create an AE event catalog in a double direct shear experiment to image the spatiotemporal evolution of microslips in the synthetic fault gouge during the stick–slip cycle. Ma et al. ( 2022 ) simulated the stick–slip behavior of a sheared granular system using a discrete element method. They investigated the relationship between macroscopic stress changes and microslips between particles using the gradient boosting method (XGBoost) and performed feature importance analysis, revealing that the local spatial autocorrelation of microslips is the most important parameter. Chaipornkaew et al. ( 2022 ) established an ML-based method to evaluate kinematic efficiency (ratio of fault slip to regional deformation), an indicator of off-fault deformation. They conducted an analog experiment in which the temporal evolution of a strike-slip fault could be tracked and trained a CNN model based on the obtained fault maps. By applying the resultant model to actual fault maps, they obtained kinematic efficiency estimates that overlap with the geologic estimation, indicating that models trained on the experimental data can quantify natural off-fault deformations.

Risk evaluation of induced earthquakes

ML is actively used to analyze human-induced seismicity caused by resource extraction, water injection, hydraulic fracturing, and geothermal development. In these cases, many microearthquakes that occur within a short period should be analyzed. ML-based techniques help to develop the microearthquake catalogs (Park et al. 2020 ; Wang et al. 2020a ; Wong et al. 2021 ; Zhou et al. 2021 ; Glasgow et al. 2021 ; Zhang et al. 2022a ; Kemna et al. 2022 ), and focal mechanism catalogs (Qin et al. 2022b ) to investigate induced seismicity.

Another application of ML in studying human-induced earthquakes is its use in searching parameters that strongly affect the resultant seismicity. For example, Pawley et al. ( 2018 ), Wozniakowska and Eaton ( 2020 ), and Hicks et al. ( 2021 ) investigated the geological parameters, operational parameters, and tectonic properties that govern the activation potential of seismicity induced by hydraulic fracturing and water injection, using supervised learning based on logistic regression. Wang et al. ( 2022a ) and Kemna et al. ( 2022 ) used a tree-based ML approach to identify critical factors (including operational parameters of hydraulic fracturing) controlling induced earthquakes. Larson et al. ( 2021 ) used classification/regression trees and neural networks for a similar problem. Szafranski and Duan ( 2022 ) trained ML models of random forest, bagging, and k-neighbors regression algorithms using numerical simulation results and used the resultant model for Bayesian inversion analysis to estimate subsurface conditions.

In addition to the important feature analysis, attempts have been made to construct ML-based models for estimating seismic potential or predicting event rates. For example, Qin et al. ( 2022a ) used a random forest model, Jakubowski and Tajduś ( 2014 ) used boosted regression trees and neural networks, and Limbeck et al. ( 2021 ) used random forest and support vector machine to construct models for predicting the rate of induced earthquakes. Picozzi and Iaccarino ( 2021 ) designed an RNN model to detect precursors of M ~ 4 earthquakes in a geothermal field. Nustes Andrade and van der Baan ( 2021 ) developed a CNN model for the real-time prediction of microearthquake activities induced by hydraulic fracturing.

Other applications in seismicity analysis

Approaches based on unsupervised clustering have a long history in seismicity analysis, and many examples have been reported. While DBSCAN (Ester et al. 1996 ) and k-means (MacQueen 1967 ) are popular techniques, other algorithms, such as HDBSCAN (Campello et al. 2013 ), OPTICS (Ankerst et al. 1999 ), Gaussian mixture models (McKean et al. 2019 ), and self-organizing maps (Ida and Ishida 2022 ) have also been used for various purposes. For example, clustering based on source parameters, such as occurrence times, hypocenters, and focal mechanisms, has been performed (Ansari et al. 2009 ; Cesca et al. 2013 ; Schoenball and Ellsworth 2017 ; Cesca 2020 ; Lurka 2021 ; Piegari et al. 2022 ). These techniques are also used in fault and fracture imaging based on hypocenter distribution (Ouillon et al. 2008 ; Ouillon and Sornette 2011 ; Wang et al. 2013 ; Custódio et al. 2016 ; McKean et al. 2019 ; Petersen et al. 2020 ; Kamer et al. 2020 ; Brunsvik et al. 2021 ; Gong et al. 2022b ), tectonic tremor analysis (Wech 2021 ), seismic hazard analysis (Weatherill and Burton 2009 ; Ansari et al. 2015 ), and detection of the preparatory process of large earthquakes (Rundle and Donnellan 2020 ; Rundle et al. 2022 ). Other examples include preparatory steps for b-value analysis (Mao et al. 2022 ), identifying active seismic periods (Ida and Ishida 2022 ), and identifying aftershock sequences (Chu and Beroza 2022 ).

ML is also used to separate background seismicity from that triggered by other earthquakes, traditionally performed using the Epidemic-Type Aftershock Sequence (ETAS) model (Ogata 1988 ) or the nearest neighbor analysis (Baiesi and Paczuski 2004 ; Zaliapin et al. 2008 ). Aden-Antoniów et al. ( 2022 ) used an ML-based model trained on the synthetic seismicity generated using the ETAS model for this task. They employed the random forest model to classify activities overlapping in the spatiotemporal distance space. Trampert et al. ( 2022 ) assumed the interevent-time distribution of induced earthquakes around gas reservoirs as a mixture of Omori, gamma, and exponential distributions. By applying gradient boosting regression, they concluded that the proportion of the Omori distribution (i.e., the fraction triggered by other earthquakes) was small. Muir and Ross ( 2023 ) used a deep Gaussian process instead of the simple background seismic activity rate function in the ETAS model to improve the seismicity model.

In seismicity analyses, ML can be used as a tool to reveal nonlinear relationships that cannot be simply modeled by existing theories. For example, DeVries et al. ( 2018 ) and Karimzadeh et al. ( 2019 ) used ML to predict the spatial pattern of aftershocks, although Mignan and Broccardo ( 2019 ) pointed out that the model proposed by DeVries et al. ( 2018 ) is unlikely to achieve better performance than the conventional approaches. Dascher-Cousineau et al. ( 2023 ) developed a DL-based seismicity model based on neural temporal point processes, called the Recurrent Earthquake foreCAST (RECAST) model, as an alternative to ETAS and demonstrated its utility and advantages. Arvanitakis and Avlonitis ( 2016 ) and Arvanitakis et al. ( 2018 , 2019 ) employed supervised learning to identify asperities on plate boundaries from seismicity statistics, such as their density, recurrence intervals, and b-values.

Application in ground-motion prediction

The prediction of ground motions due to earthquakes is a major research topic in seismology and earthquake engineering; its objective is to predict the intensity or time series of ground motions caused by a given earthquake at a target location. Ground-motion prediction techniques are used for seismic hazard assessment, early warning of strong motions, and damage estimation after an earthquake. Recently, ML has been applied to this prediction and for creating supporting information. Hereafter, we review ML applications in (1) predicting ground-motion intensity from “features”, which are individual measurable properties, such as distance from an earthquake and earthquake magnitude; (2) predicting time series of ground motions from features; (3) predicting ground-motion intensity from time series input; (4) predicting ground-motion time series from time series input; (5) other related studies. The relationship among 1–4 is visualized in the matrix component of Fig.  1 .

Prediction of ground-motion intensity from features

Empirical equations called ground motion prediction equations (GMPEs) have been used to predict the intensity of ground motions (i.e., seismic intensity and peak ground acceleration) at a target location caused by a given earthquake. GMPEs have been constructed by regressing past ground-motion records, considering the geophysical model between an earthquake and ground motions. Since this is a supervised ML problem, ML regression techniques have been applied. The applications were underway before the third Artificial Intelligence (AI) boom began in the late 2010s (Kerh and Ting 2005 ; Ahmad et al. 2008 ). Recently, many studies have been conducted using various ML methods: neural networks (Derras et al. 2012 , 2014 ; Dhanya and Raghukanth 2018 ; Wang et al. 2020b ; Khosravikia and Clayton 2021 ; Okazaki et al. 2021c , b ; Ji et al. 2021 ), tree-based models such as random forest and XGBoost (Trugman and Shearer 2018 ; Kubo et al. 2020 ; Khosravikia and Clayton 2021 ; Oana et al. 2022 ), support vector regression (Tezcan and Cheng 2012 ; Khosravikia and Clayton 2021 ; Hu et al. 2022 ), Gaussian process regression (Hermkes et al. 2014 ; Alimoradi and Beck 2015 ; Lavrentiadis et al. 2023 ), and Bayesian network (Kuehn et al. 2011 ). The prediction targets of ground-motion intensity include peak ground acceleration (PGA), peak ground velocity, response spectra, and seismic intensity. While early ML models predicted a single indicator, more recent multi-output models can predict several indicators simultaneously (Dhanya and Raghukanth 2018 ; Oana et al. 2022 ; Hu et al. 2022 ).

An advantage of ML models over conventional GMPEs is that nonparametric ML methods can learn the functions of ground-motion models directly from data without assuming regression equations, resulting in a more accurate and useful predictor within the range of training data. Some studies have reported that ML prediction models have higher accuracy than the existing GMPEs (Khosravikia and Clayton 2021 ; Oana et al. 2022 ; Seo et al. 2022 ). Khosravikia and Clayton ( 2021 ) reported that when sufficient data are available, the ML prediction models provide more accurate estimates than the conventional linear regression-based method.

Another advantage of the ML model is that new data or features (explanatory variables) not used in conventional GMPEs can be incorporated. Conventional GMPEs used earthquake magnitude, distance from an earthquake (e.g., hypocentral distance, shortest distance from the fault, and Joyner–Boore distance), depth of an earthquake, and site information below the predicted point, such as Vs30 (average S-wave velocity up to a 30 m depth). The ML prediction model designed by Oana et al. ( 2022 ) adopted additional features, such as the earthquake’s location, the prediction point’s location, and back azimuth information. Their ML model with the additional features had a lower logarithmic standard deviation value than the previous GMPEs. Esteghamati et al. ( 2022 ) designed an ML predictor with information on horizontal-to-vertical spectral ratios (HVSR) as explanatory variables. Although a simple proxy, such as Vs30, has often been used in previous studies as site information, the use of more detailed information, such as HVSR, is expected to improve the accuracy of ground-motion prediction. Moreover, a ground-motion prediction model that directly accounts for the site amplification at the target station can be made by learning ground-motion data labeled for each observation station. Okazaki et al. ( 2021c ) proposed an ML model that uses a one-hot representation of site ID, which achieved accurate prediction considering site-specific ground-motion amplification and avoided overfitting at sites with few records.

Furthermore, ML leads to a more flexible prediction. Although previous studies made predictions at individual points, Lilienkamp et al. ( 2022 ) achieved “areal” prediction by treating the explanatory and predictive variables as map-shaped data. Rekoske et al. ( 2023 ) presented a reduced-order modeling approach based on interpolated proper orthogonal decomposition to predict the spatial distribution of peak ground velocity. Sreenath and Raghukanth ( 2023 ) constructed a ground-motion model based on the Bayesian Neural Network (BNN) (Jospin et al. 2022 ), in which the Bayesian framework is incorporated into neural networks. Their approach could enable the evaluation of uncertainties in the prediction of earthquake motion because BNN facilitates probabilistic prediction.

To mitigate the negative effect of the black-box nature of AI models, eXplainable AI (XAI) techniques have recently been developed to help humans understand the reasons behind decisions or predictions made using AI models. The applications of the XAI techniques are useful in understanding the behavior of an ML-based ground-motion prediction model. Mohammadi et al. ( 2023 ) applied Shapley additive explanation (SHAP), which was developed based on the game theory to explain individual predictions of ML models (Lundberg and Lee 2017 ), to explore the importance of each explainable variable in the XGBoost prediction model.

Prediction of ground-motion time series from features

Numerical simulations have often been used to obtain seismic waveforms at a target location due to a given earthquake. With the development of observation networks, the accumulation of ground-motion records, advancements in the subsurface structure, refinement of numerical simulation techniques for ground motions, and increased machine power, numerical simulations of three-dimensional seismic wavefield have been conducted by assuming a subsurface velocity structure model that reflects the real three-dimensional heterogeneity (Moczo et al. 2014 ; Igel 2017 ).

Recently, researchers have explored a data-driven approach using generative ML models for obtaining seismic waveforms based on past ground-motion records. Generative ML models are a class of statistical models that attempt to capture the underlying probability distribution of data and can be trained to generate data instances (Bengio et al. 2013 ; Jebara 2012 ). One of the major generative models is GAN (Goodfellow et al. 2014 ), in which two networks are trained sequentially: a generator network to produce pseudo data and a discriminator network to distinguish between the truth or falsity of data. The well-trained generator can be used as a generative ML model. Mirza and Osindero ( 2014 ) proposed a conditional GAN (CGAN) that controls the generation process by considering additional features. Combining CGAN and Wasserstein GAN (Arjovsky et al. 2017 ) with the training of a massive set of strong motion recordings in the time domain, Florez et al. ( 2022 ) developed a generative model that synthesizes realistic three-component accelerograms conditioned on magnitude, distance, and Vs30. Their generative model captures the most relevant statistical features of the acceleration spectra and waveform envelopes. The output seismograms display clear P‐ and S‐wave arrivals with the appropriate energy content and relative onset timing. Moreover, Esfahani et al. ( 2023 ) developed a generative model for simulating ground-motion recordings, which combines a CGAN to predict the amplitude part of the time–frequency representation of ground-motion records and a phase-retrieval method. Matsumoto et al. ( 2023 ) constructed a ground motion generation model using GANs and evaluated its performance. They demonstrated through numerical experiments that their proposed model is probabilistic, approximates the probabilistic distribution of the dataset of observed records, and generates realistic ground-motion time histories with various characteristics in the time and frequency domains. As a new-generation model, diffusion models (Ho et al. 2020 ; Rombach et al. 2021 ) have recently attracted attention and will be used for ground-motion prediction. The data-generation process with a generative ML model involves randomness, and the generated data instances differ slightly even if they are generated under a common condition. How to use the generated instances with randomness remains unclear. Moreover, strategies to evaluate the validity of the generated instances are needed.

Another possible ML approach for obtaining ground-motion time series is PINNs, which can consider physical laws. Several studies have applied PINNs to the forward simulation of seismic waveforms, including in the simulation of acoustic wavefield described by a frequency-domain wave equation (Alkhalifah et al. 2020 ; Song et al. 2021a , b ; Huang et al. 2021 ; Huang and Alkhalifah 2022 ; Song and Wang 2022a ), the reconstruction inversion of acoustic wavefields (Song and Alkhalifah 2022 ), and the simulation of two-dimensional (2D) SH wave propagation (Ding et al. 2023a , b ). Furthermore, Rasht-Behesht et al. ( 2022 ) proposed an approach to solve wave propagation and full waveform inversion based on PINNs. Moreover, neural operators, which learn mapping between function spaces via supervised training to solve PDEs (Kovachki et al. 2021 ), have attracted attention in recent years and have been applied to the construction of a surrogate model for wave propagation simulation (Yang et al. 2021b , 2023 ; Song and Wang 2022b ; Zhang et al. 2023 ; Zou et al. 2023 ; Lehmann et al. 2024 ). Although PINNs and neural operators remain in methodological development and have limited applications, further development in ground-motion prediction is expected.

Prediction of ground-motion intensity from time series

Onsite prediction of the final ground-motion intensity at a target site using the first few seconds of its observations is a major topic of EEW (Nakamura 1988 ; Wu and Kanamori 2005 ; Spallarossa et al. 2018 ). Recently, this has been approached using ML techniques, such as random forest (Hu et al. 2023 ), CNN (Jozinović et al. 2020 ; Zhang et al. 2022b ), support vector machine (Song et al. 2022 ), Graph Neural Network (Bloemheuvel et al. 2022 ), and LSTM (Wang et al. 2022b , 2023a ). Additionally, ML has been applied to discriminate in real-time whether a signal stems from an actual earthquake based on the initial part of ground motions (Li et al. 2018 ; Meier et al. 2019 ; Liu et al. 2022 ). Although these studies adopted a single-station approach where inputs and outputs were closed at one station, Münchmeyer et al. ( 2020 ) proposed a multiple-station approach that predicts final PGA values at multiple target stations using initial raw records at multiple input stations.

Prediction of ground-motion time series from time series

Recently, predicting future or present ground motions directly from currently observed ground motions, skipping source estimation, has been actively studied in the research field of EEW (Kodera et al. 2018 ; Hoshiba 2021 ), and ML has been applied. Otake et al. ( 2020 ) proposed a DL model to predict the present seismic intensity at the target station based only on the observation records at the surrounding stations, demonstrating that its prediction performance is better than adopting the maximum or weighted average of the input data. Additionally, Fornasari et al. ( 2022 ) developed an ensemble ML model combining CNN and Voronoi tessellation, which could reconstruct the ground-shaking field in real time based on the point-cloud spatial distribution of current ground-shaking data recorded at each observation station. Datta et al. ( 2022 ) designed a deep spatiotemporal RNN to predict future intensity time series at multiple stations directly from their current ground-motion observations. Tamhidi et al. ( 2022 ) proposed an approach where ground-motion time series at target sites are constructed from a set of observed motions using a Gaussian process regression, which treats the real and imaginary parts of the Fourier spectrum as random Gaussian variables. Furumura and Oishi ( 2023 ) developed a DL approach for early prediction of long-period ground motions at a target point far from the earthquake source based on waveform observations near the source. This approach could effectively forecast long-period ground motions of large earthquakes regarding amplitude, waveform envelope shape, spectral characteristics, and duration.

Due to computational constraints and insufficient knowledge of the underground structure at short wavelengths, physics-based simulation (PBS) ground motions are reliable only in the long-period range (typically, approximately 1 s). ML has been applied to convert long-period to short-period PBS waveforms and obtain broadband ground-motion waveforms. In the approach proposed by Paolucci et al. ( 2018 ), the short-period response spectra were estimated using an ML model from the long-period ones simulated by PBS. Then, the short-period components of the original PBS waveforms were enriched using a stochastic simulation to match the estimated response spectra. Sharma et al. ( 2022 ) upgraded the estimation of the short-period response spectra proposed by Paolucci et al. ( 2018 ) in a way that not only response spectra of long-period PBS waveforms but also additional features, such as source, path, and site parameters, were used as explainable variables. Furthermore, Okazaki et al. ( 2021a ) explored an approach that generates consistent broadband waveforms using past observation records. Long-period Fourier amplitude spectra and acceleration envelopes were transformed to short-period ones using a neural network and an embedded space obtained by the nonlinear dimensionality reduction of t-SNE (Van der Maaten and Hinton 2008 ), respectively; they were combined to produce a broadband waveform. The notable features of their study include (1) using the Wasserstein distance in the dimensionality reduction of envelopes to capture their global properties and (2) simultaneously embedding pairs of long- and short-period envelopes into a common space to obtain the relationship from a limited amount of ground-motion data.

The denoising of seismic waveform records is related to the issue discussed in this section. Several studies have been conducted to extract the target signal from originals containing waves from various sources or separate signals and noise (Zhu et al. 2019b ; Tibi et al. 2021 , 2022 ; Dalai et al. 2021 ; Novoselov et al. 2022 ; Yin et al. 2022a ; Xu et al. 2022 ; Wang and Zhang 2023 ). These efforts will lead to the effective use of observation data unanalyzed due to a low signal-to-noise ratio.

Studies related to ground-motion prediction

ML has been applied to estimate site amplification characteristics, of which frequency dependency varies among sites. Roten and Olsen ( 2021 ) developed a DL model to estimate the relative site amplification characteristics between borehole and surface seismograms from discretized S-wave velocity profiles. Pan et al. ( 2022 ) used a DL model to estimate S-wave site amplification characteristics from the information on microtremor HVSR. Zhu et al. ( 2023b ) proposed an ML model to obtain site amplification factors from site proxies and suggested that the ML model performs better than the conventional approach. Zhu et al. ( 2023a ) trained a supervised DL model to recognize and separate seismic site response from single-station seismograms. Pilz et al. ( 2020 ) proposed a data-driven identification method based on information from various sites to identify a reference site.

Furthermore, ML techniques have been applied to seismic hazard assessment. Numerical simulations of ground motions based on many earthquake scenarios should be conducted and sampled to quantify variations in seismic hazard assessment. Imai et al. ( 2021 ) proposed a method to generate many scenario earthquake shaking maps from existing shaking maps using modal decomposition of proper orthogonal decomposition and an empirical copula.

The quality assessment and selection of ground-motion records are necessary to improve the ground-motion prediction model. Dupuis et al. ( 2023 ) applied a supervised DL-based model to this issue. Their model can estimate the quality and minimum usable frequency for each record component. It can handle one-, two-, or three-component records, providing flexibility to assess record quality based on various requirements.

Tsunamis are another phenomenon caused by earthquakes besides ground shaking. Several studies have applied ML to tsunami prediction (Fauzi and Mizutani 2020 ; Makinoshima et al. 2021 ; Liu et al. 2021 ; Kamiya et al. 2022 ; Rim et al. 2022 ).

Approach to imbalanced ground-motion data

One significant problem with applying ML in ground-motion prediction is the imbalance in observation data of ground motions. Observation data in the geophysical field are often imbalanced. In the case of ground-motion data, few records of strong shaking, near source faults, and large-magnitude earthquakes are available. Figure  3 shows the distribution of the ground-motion dataset used by Kubo et al. ( 2020 ): ground-motion records in the range of 1–20 cm/s/s account for much of the dataset, whereas there are few records of strong motions over 1,000 cm/s/s. Moreover, the spatial distribution of earthquakes and their focal mechanisms are strongly biased. Most observations of ground motions have been made on land. Learning such biased data may have undesirable effects on the ML prediction model. For example, Kubo et al. ( 2020 ) showed that ML training with unbalanced ground-motion data leads to biased prediction.

Distribution of ground-motion dataset used by Kubo et al. ( 2020 ) with the heat map showing the relationship between epicentral distance and PGA with their histograms (Fig. 5 of Kubo et al. ( 2020 )). They used 186,310 ground-motion records in Japan, which were recorded from 1997 to 2015 by K-NET and KiK-net of National Research Institute for Earth Science and Disaster Resilience. The heat-map color denotes the number of ground-motion records

One approach to solving this issue is augmentation with simulated data. Withers et al. ( 2020 ) used a database of synthetic ground motions extracted from the Southern California CyberShake study to build an ML prediction model of ground-motion intensities in Southern California. This model had behavior and performance similar to empirically based models. Raghucharan et al. ( 2021 ) appended the simulated data from the validated seismological model to the observation data so that the trained dataset for their ML model covered a wide range of magnitude and distance, filling the data gap region. Lehmann et al. ( 2023 ) prepared a synthetic database based on many PBSs considering various source mechanisms or finite faults for constructing an ML model to simulate low-frequency ground-motion parameters for arbitrary focal mechanisms or finite fault sources. ML training with simulated data alone has been done in research fields where observation data are fundamentally scarce, such as in the cases of tsunamis (Makinoshima et al. 2021 ) and gravity waves (Licciardi et al. 2022 ). The emulation of ground motions using ML surrogate models, such as generative ML models, PINNs, and neural operators, is expected to be used for data augmentation.

Another approach to address this issue is incorporating geophysical findings into the ML model. Several studies have incorporated conventional GMPEs into the ML model. GMPEs are stable for extrapolation or low data density because conventional GMPEs are based on the geophysical background of ground motions. Khosravikia and Clayton ( 2021 ) reported that the approach of conventional GMPEs is better than ML-based models when training data are limited. Kubo et al. ( 2020 ) proposed a hybrid approach of ML and conventional GMPE for predicting ground-motion intensity. In this approach, an ML model was trained with the residuals between observations and predictions using GMPE, and the sum of the predictions using GMPE and the ML model was used as the prediction of the hybrid model. They demonstrated that this hybrid approach reduces the underestimation of strong motions found in the prediction model using only ML and performs better than the individual approaches. Oana et al. ( 2022 ) included the predicted value of GMPE in the explainable variable of the ML prediction model.

Direct incorporation of geophysical knowledge into an ML model has been conducted. Previous studies on ground motions clarified the relationship between ground motions and related factors. Generally, ground motions increase with magnitude and decrease with distance, and they tend to decrease with Vs30 and increase with the thickness of sediment layers. To follow these relationships in an ML prediction model, Okazaki et al. ( 2021b ) proposed a neural network model with a monotonic dependence on the input variables. They demonstrated that this approach reduces overfitting for training data and improves generalized performance.

Transfer learning (Pan and Yang 2010 ; Bozinovski 2020 ) can be used to address ground-motion prediction problems. Here, an existing ML model trained with ground-motion records in areas with large amounts of data is used as an initial model to develop an ML model in areas with few records. Jozinović et al. ( 2022 ) introduced transfer learning to develop an ML model of onsite EEW for a small dataset; their approach improved the prediction performance regarding outliers, bias, and variability of the residuals between predicted and actual values.

As described in this section, many studies have applied ML to earthquake ground-motion prediction, resulting in significant progress, including reporting predictions that are more accurate than existing models and realizing predictions that were previously difficult to predict. However, ML-based ground motion prediction is not widely used in practical applications such as earthquake hazard assessment and earthquake early warning. The black-box nature of ML is likely one factor limiting its use. Further research is needed to evaluate the limits of applicability of ML predictions and their uncertainty, versatility, and reproducibility.

While the ground-motion prediction model will continue to become more sophisticated, it is also essential to update the ground-motion data. To achieve this, it is crucial to maintain and upgrade strong-motion observation networks and to steadily increase observation records. Furthermore, the usability, accessibility, and transparency of the observation records must be continuously improved to encourage the further use of ground-motion data. One such effort is the construction and publication of a “flat file”, a data set that integrates information on ground-motion intensity measurements, seismic source information, and site information, such as the NGA-WEST2 database (Ancheta et al. 2014 ) and the New Zealand strong motion database (Van Houtte et al. 2017 ). As mentioned in the previous subsection, extending the dataset based on PBSs is also important. The simulations can create records for huge earthquakes or those near faults, for which there are few observation records. This will help compensate for the imbalance in the observation-based ground-motion dataset. Such efforts are useful in applying ML because the imbalance in training data directly leads to the reduction of the prediction performance of ML models. Data emulation with ML surrogate models, such as generative ML models, PINNs, and neural operators, will also be used.

Several studies have raised issues regarding the effect of ergodic assumption in ground-motion prediction models on their prediction results (Anderson and Brune 1999 ; Abrahamson et al. 2019 ; Lavrentiadis et al. 2023 ). The ergodic assumption is that the median and aleatory variability of a ground-motion prediction model are applicable to any location within the broad tectonic category. Although most ground-motion prediction models are based on this assumption, the ergodic ground-motion prediction models may not work well for a specific site/source location because of significant systematic differences in ground motion depending on the location of the site and source. Therefore, a growing need exists for developing non-ergodic ground-motion prediction models that explicitly model the location-specific effects, which can lead to a reduction in aleatory variability. The application of ML will prove essential in the construction of site-specific or location-specific ground-motion prediction models because the situation is more conducive to the flexibility of ML, which can handle various explanatory variables.

Bergen et al. ( 2019 ) indicated that ML use in solid earth geoscience can be broadly divided into three categories: automation, modeling (simulation and inversion), and discovery. For example, phase picking is an example of automation, while ground motion prediction falls under modeling. The current use of ML in seismology is primarily focused on the automation category. Applying ML to modeling or discovery may be more difficult; however, we look forward to further research in this area.

Application to geodetic data

This section summarizes the application of ML to geodetic data. We primarily focus on topics related to crustal deformation caused by tectonic motion.

Clustering analysis

GNSS observations record seismic and aseismic signals as well as the long-term secular motion of the Earth’s surface. These long-term secular velocity fields contain information regarding rigid block motion, internal deformation, and strain accumulation processes due to plate motions, such as subduction, opening, and collision. Data-driven approaches based on clustering algorithms have been developed to identify regional differences in these velocity fields. Clustering approaches can objectively determine block boundaries without prior information, such as fault location, geological information, or underlying deformation processes. These algorithms can be broadly divided into hard and soft clustering. Hard clustering algorithms determine the cluster to which each observation belongs, and soft clustering allows each station to belong to multiple clusters. In other words, the soft clustering approach provides information on the uncertainty of each cluster at each station.

Two approaches of hard clustering algorithms have been applied to GNSS velocity fields. The first approach is hierarchical clustering, which initiates the clustering procedure with each cluster comprising a single data point. Then, two clusters are sequentially combined based on the similarity in their data. Consequently, all data can be categorized into a hierarchical tree or a dendrogram. When focusing on a small number of clusters (i.e., higher levels in the dendrogram), each boundary represents a large-scale difference in GNSS velocities, whereas many clusters with lower-level clustering correspond to small-scale boundaries. The advantage of this approach is that the clustering result is uniquely determined, and there is no need to define the number of clusters. Once two stations are connected in the same cluster at a lower level, these two stations will never be categorized into different clusters at a higher level.

The second approach is the partitioning-optimization approach, in which the number of clusters is first determined, followed by clustering using a measure of similarity between observations. The clustering results depend on the initial clusters, which must be determined beforehand. The robustness of these initial clusters can be examined by conducting clustering with different sets of initial clusters. However, achieving a global minimum solution is not guaranteed.

Simpson et al. ( 2012 ) proposed a hierarchical agglomerative clustering approach using the similarity of velocity vectors as a metric for clustering. They applied it to the GNSS data in the San Francisco Bay Region, California, where the tectonic setting is simple and dominated by a strike-slip motion. Additionally, they proposed the k-means method as a partitioning-optimization clustering approach and compared it with hierarchical agglomerative clustering. Both methods yielded substantially different clustering results depending on the number of clusters; however, the geographical clustering maps were roughly similar. This approach assumes a flat earth and is, thus, suitable for velocity fields that can be approximated using the translational motion of blocks, such as in the San Francisco Bay Region. In other words, this approach is useful when the Euler pole is located at a great distance. Following this study, Savage and Simpson ( 2013a ) generalized the k-means approach proposed by Simpson et al. ( 2012 ) to be applicable to a spherical earth. For this purpose, they proposed the Euler-Vector clustering algorithm to minimize the residuals between the observed velocity and calculated velocity obtained using the Euler-Vector in the cluster to which each station belongs. Starting from the initial grouping obtained using the k-means clustering, the Euler-Vector clustering was applied to the Mojave Block, southeastern California (Savage and Simpson 2013a ) and the California–Nevada region (Savage and Simpson 2013b ), yielding a clustering result similar to the k-means clustering. The velocity fields targeted by these previous studies are oriented in roughly similar directions with different amplitudes, indicating that the Euler poles are at distant locations.

Savage and Wells ( 2015 ) applied a similar clustering approach across a broad area in the Pacific Northwest, USA, from 38°N to 49°N, including California, Nevada, Oregon, and Washington. The velocity field in this target area is approximated by a single Euler Pole, indicating that the velocity field is rotating clockwise and the Euler Pole is close to the target area. Although this velocity field incorporates information on strain accumulation due to plate subduction in Cascadia and internal deformation (or includes noise for clustering), they identified blocks that align with geologically determined boundaries. Subsequent studies applied this hard clustering method to the GNSS velocity fields in the southern San Andreas Fault (Thatcher et al. 2016 ), southwest Japan (Savage 2018 ), Taiwan (Takahashi et al. 2019 ), Turkey (Özdemir and Karslıoğlu 2019 ), and New Zealand (Takahashi and Hashimoto 2022 ), and discussed regional characteristics of block boundaries.

As illustrated above, GNSS velocity is a powerful tool for objectively identifying crustal structures. However, the observed velocities are often contaminated by various types of noise, such as atmospheric and ionospheric delays, the local motion of observation sites, and orbit errors. Takahashi et al. ( 2019 ) proposed a method of assessing stability based on information entropy to evaluate the robustness of the clustering results obtained using such a noisy dataset. Numerous synthetic datasets were prepared to calculate the information entropy by adding noise and conducting clustering for each dataset. Then, for each pair of stations, the information entropy was defined so that it is minimized when the two vectors are classified into the same or different clusters. That is, when these two data points are randomly classified, the information entropy reaches its maximum. By summing the information entropy for all pairs at each station, the entropy or robustness of the clustering result at each station was obtained (see Fig. 8 in Takahashi et al. ( 2019 ) and Fig. 6 in Takahashi and Hashimoto ( 2022 )).

In most previous studies, clustering analysis was performed based on velocity fields; however, Yáñez-Cuadra et al. ( 2023 ) employed strain rate invariants and rotation rate fields, derived from velocity fields. Applying the hierarchical agglomerative clustering algorithm to the GNSS data in the Chilean subduction zone revealed a correlation between clustering results and the segmentation of the seismogenic zone and regional geological structures.

Unlike hard clustering, the soft clustering approach provides a probability of the cluster to which each station belongs. This approach directly identifies stations or regions with ambiguous crustal structure boundaries without additional evaluation of the clustering result. Özdemir and Karslıoğlu ( 2019 ) compared three hard clustering methods including k-means, hierarchical agglomerative clustering, and Gaussian mixture model with a soft clustering version of Gaussian mixture model fits. All the hard clustering algorithms successfully clustered GPS velocities in Turkey into the Eurasian and Arabian blocks, separated by the Anatolian strike-slip faults. However, additional clustering results within the Anatolian block varied depending on the chosen clustering algorithms. This ambiguity is clearly illustrated by the soft clustering method. Mitsui and Watanabe ( 2020 ) developed a different soft clustering approach using the fuzzy c-means method. They discussed the tectonics in the Izu Peninsula, a collisional zone with a complex deformation field in Japan.

Granat et al. ( 2021 ) developed an open-source code for clustering GNSS velocities using methods available in the scikit-learn package in Python. The clustering results can be easily displayed via Google Earth or Google Maps using this code.

Detection of geodetic signals due to seismic/aseismic phenomena

ML techniques are powerful tools for extracting spatiotemporal signals of interest. A primary focus in recent geodetic studies is the detection of transient signals, such as slow slip events (SSEs). Accurately detecting such transient signals contributes to our understanding of slow earthquake activities while providing an accurate estimation of interseismic velocities that reflect the state of interplate coupling. Signals due to SSEs are often so weak that they can be buried in noise; therefore, methods that can effectively detect these small signals are required.

Several approaches have been proposed to detect SSEs: 1) extraction solely from geodetic data (Nishimura et al. 2013 ; Crowell et al. 2016 ; Yano and Kano 2022 ); 2) extraction using templates (Riel et al. 2014 ; Rousset et al. 2017 ; Okada et al. 2022 ); and 3) using other slow earthquake phenomena as a reference (Frank et al. 2015 ; Bartlow 2020 ). Additionally, ML-based approaches have been suggested. Most proposed methods adopted a supervised approach to detection or classification problems.

He et al. ( 2020 ) combined RNNs and CNNs to detect SSEs, applying the Ocean Bottom Pressure (OBP) data obtained in New Zealand. The input for their model is a time series, and the output is a label indicating whether SSEs occur in the middle of the time series. They demonstrated that the proposed method performed better than the matched filter approach. Because seafloor geodetic data, such as OBP, are typically noisier than onland GNSS data, successful extraction of SSE signals will improve the understanding of slip behavior, especially in the shallower parts of subduction zones.

Rouet-Leduc et al. ( 2021 ) applied a deep autoencoder to a time series of Interferometric Synthetic Aperture Radar (InSAR) images to automatically detect millimeter-scale crustal deformation without fault location information. Their method was designed to mitigate noise in the InSAR time series, leveraging the characteristic that atmospheric errors do not correlate over a day and tectonic signals result in permanent deformation. They detected crustal deformation caused by a slow earthquake in the Anatolian Fault, Turkey, which was larger than previously considered. Xue and Freymueller ( 2023 ) proposed an RNN-based approach to obtain a time series of the probability of transient signals being included in the GNSS time series using a single station. Their model sequentially refers to each data only once; thereby, efficiently obtaining output probabilities compared with common methods using sliding windows. After training their model using synthetic data, they applied it to detect the SSEs in Cascadia between 2005 and 2016. Their detection results were consistent with the previously identified characteristics of SSEs such as timings, durations, and areas. Furthermore, they demonstrated that the proposed method is robust despite gaps in the time series. Based on GNSS data, Costantino et al. ( 2023 ) developed DL methods for estimating seismic and aseismic source characteristics. By, respectively, training spatial, temporal, and spatiotemporal synthetic datasets in DL models, the spatiotemporal data with the transformer displayed the best performance in estimating source parameters of earthquakes along the Japan Trench in northeast Japan. They highlighted the potential of geodetic data to improve a low detection limit using an ML approach. The spatial pattern of crustal deformation is useful for the real-time detection of large earthquake signals because of its usage of unsaturated displacement waveforms. Lin et al. ( 2021 ) proposed the GNSS-based EEW algorithm, M-LARGE, and demonstrated that the proposed method outperformed other GNSS-based non-DL methods.

Modeling and prediction of crustal deformation

AI techniques have been introduced to model crustal deformation. DeVries et al. ( 2017 ) proposed a viscoelastic modeling method using a neural network trained on theoretical modeling results. By introducing neural networks, the calculation can be drastically accelerated compared with the forward calculation. Okazaki et al. ( 2022 ) proposed a PINN approach to model crustal static deformation due to an earthquake. By using a polar coordinate system, the displacement discontinuity on a fault can be accurately modeled as a boundary condition. Fukushima et al. ( 2023 ) used a PINN approach for earthquake cycle simulation that calculates the temporal evolution of slip velocities of SSEs using a spring-slider system. They additionally proposed an alternative inversion approach based on PINNs for frictional parameter estimation. Due to its flexibility, PINNs can readily incorporate complex models such as heterogeneous underground structures and three-dimensional plate geometries. Therefore, PINNs will be a powerful tool for geodetic inversion. Another application that integrates geodetic data and ML techniques is the prediction of time series. Fukushima et al. ( 2023 ) proposed a PINN-based scheme for predicting the short-term evolution of slip velocities along the subducting plate interface.

ML has been adopted to predict postseismic deformation. Time series of postseismic deformation are often modeled as a superposition of logarithmic and exponential functions (Tobita 2016 ). Yamaga and Mitsui ( 2019 ) applied an RNN to this problem and predicted future time series following the 2011 Tohoku-oki earthquake. Their RNN inputs a 1-year GNSS time series and forecasts displacement on the day following the input time series. After training the RNN model using GNSS stations in northeast Japan, they conducted predictions of GNSS time series that were not used in training. Compared with the traditional method of fitting logarithmic and exponential functions, the RNN model had better predictive abilities with smaller residuals between the predicted and observed displacements. They showed that the residual time series of the RNN outputs represented the dominant mechanisms depending on the elapsed time from the mainshock, that is, the afterslip decay and, consequently, the domination of viscoelasticity.

As described in this section, ML techniques have been introduced to geodetic data analysis such as clustering, detection of events, and modeling. Nonetheless, fewer application examples have been reported compared with seismic data analysis as reviewed in this paper. Clustering analysis is one of the well-studied topics. The previous studies have demonstrated that the proposed algorithms are widely applicable to crustal deformation in various tectonic regions. The resulting clusters provide boundaries of crustal structures with their uncertainties from regional to global scales, contributing to our understanding of the tectonics in the target areas. In contrast, in recent years, the detection of tectonic signals, modeling, and prediction of crustal deformation are emerging topics of ML application to geodetic data. As new ML-based approaches are continuously proposed, they would detect a greater number of hidden events, discover new aseismic phenomena, or be utilized as a tool that can easily incorporate complex geometries or spatial heterogeneity of underground structures for geodetic modeling. These achievements will eventually contribute to further understanding of the physical processes associated with the whole earthquake cycle.

Conclusions

In this paper, we reviewed ML applications in several fields of earthquake seismology, especially in the development of earthquake catalog, seismicity analysis, ground motion prediction, and application to geodetic data. ML technologies have significantly advanced these fields; however, unique challenges persist. For example, the imbalance in natural datasets is problematic in many cases, possibly causing misevaluation or misinterpretation. Effective approaches to address this problem include data augmentation, simultaneous use of domain knowledge, and transfer learning. Although challenges arise from the black-box nature of DL, the latest techniques, such as PINNs, neural operators, BNN, and XAI, can address them. The efficiency, accuracy, and flexibility of ML are the driving forces behind the establishment of its usage for various tasks in earthquake seismology. There remain many problems where ML can effectively solve, and its application will further expand and advance our knowledge of earthquake seismology.

Availability of data and materials

Not applicable.

Abbreviations

Acoustic emission

Artificial intelligence

Bayesian neural network

Conditional generative adversarial network

• Deep learning

Earthquake early warning

Epidemic-type aftershock sequence

Generative adversarial network

GNSS Earth Observation Network System

Global Navigation Satellite System

Ground motion prediction equation

Graphics processing unit

Horizontal-to-vertical spectral ratios

Interferometric synthetic aperture radar

Locally sensitive hashing

Long-term average

• Machine learning

Ocean bottom pressure

Physics-based simulation

Partial differential equation

Peak ground acceleration

Physics informed neural networks

Recurrent neural network

Short-term average

Slow slip event

EXplainable AI

Abrahamson NA, Kuehn NM, Walling M, Landwehr N (2019) Probabilistic seismic hazard analysis in California using nonergodic ground-motion models. Bull Seismol Soc Am 109:1235–1249. https://doi.org/10.1785/0120190030

Article   Google Scholar  

Aden-Antoniów F, Frank WB, Seydoux L (2022) An adaptable random forest model for the declustering of earthquake catalogs. J Geophys Res 127:e2021JB023254. https://doi.org/10.1029/2021jb023254

Article   ADS   Google Scholar  

Agata R, Shiraishi K, Fujie G (2023) Bayesian seismic tomography based on velocity-space Stein variational gradient descent for physics-informed neural network. IEEE Trans Geosci Remote Sens 61:1–17. https://doi.org/10.1109/tgrs.2023.3295414

Ahmad I, El Naggar MH, Khan AN (2008) Neural network based attenuation of strong motion peaks in Europe. J Earthquake Eng 12:663–680. https://doi.org/10.1080/13632460701758570

Albert S, Linville L (2020) Benchmarking current and emerging approaches to infrasound signal classification. Seismol Res Lett 91:921–929. https://doi.org/10.1785/0220190116

Ali A, Sheng-Chang C, Ali SH (2022) Integration of density-based spatial clustering with noise and continuous wavelet transform for feature extraction from seismic data. Pure Appl Geophys 179:1183–1195. https://doi.org/10.1007/s00024-022-02980-7

Alimoradi A, Beck JL (2015) Machine-learning methods for earthquake ground motion analysis and simulation. J Eng Mech 141:04014147. https://doi.org/10.1061/(ASCE)EM.1943-7889.0000869

Alkhalifah T, Song C, Waheed UB (2020) Machine learned Green’s functions that approximately satisfy the wave equation. In: SEG Technical Program Expanded Abstracts 2020. Society of Exploration Geophysicists

Allen RV (1978) Automatic earthquake recognition and timing from single traces. Bull Seismol Soc Am 68:1521–1532. https://doi.org/10.1785/BSSA0680051521

Ancheta TD, Darragh RB, Stewart JP et al (2014) NGA-West2 database. Earthq Spectra 30:989–1005. https://doi.org/10.1193/070913EQS197M

Anderson JG, Brune JN (1999) Probabilistic seismic hazard analysis without the ergodic assumption. Seismol Res Lett 70:19–28. https://doi.org/10.1785/gssrl.70.1.19

Anikiev D, Waheed UB, Staněk F et al (2022) Traveltime-based microseismic event location using artificial neural network. Front Earth Sci 10:1046258. https://doi.org/10.3389/feart.2022.1046258

Anikiev D, Birnie C, Waheed UB et al (2023) Machine learning in microseismic monitoring. Earth-Sci Rev 239:104371. https://doi.org/10.1016/j.earscirev.2023.104371

Ankerst M, Breunig MM, Kriegel H-P, Sander J (1999) OPTICS: Ordering points to identify the clustering structure. SIGMOD Rec 28:49–60. https://doi.org/10.1145/304181.304187

Ansari A, Noorzad A, Zafarani H (2009) Clustering analysis of the seismic catalog of Iran. Comput Geosci 35:475–486. https://doi.org/10.1016/j.cageo.2008.01.010

Ansari A, Firuzi E, Etemadsaeed L (2015) Delineation of seismic sources in probabilistic seismic-hazard analysis using fuzzy cluster analysis and Monte Carlo simulation. Bull Seismol Soc Am 105:2174–2191. https://doi.org/10.1785/0120140256

Arjovsky M, Chintala S, Bottou L (2017) Wasserstein generative adversarial networks. In: Precup D, Teh YW (eds) Proceedings of the 34th International Conference on Machine Learning. PMLR, pp 214–223

Arvanitakis K, Avlonitis M (2016) Identifying asperity patterns via machine learning algorithms. IFIP advances in information and communication technology. Springer International Publishing, Cham, pp 87–93

Google Scholar  

Arvanitakis K, Avlonitis M, Papadimitriou E (2018) Introducing stochastic recurrence interval to classification algorithms for identifying asperity patterns. Physica A 512:566–577. https://doi.org/10.1016/j.physa.2018.08.142

Arvanitakis K, Karydis I, Kermanidis KL, Avlonitis M (2019) A machine learning approach for asperities’ location identification. Evol Syst 10:41–50. https://doi.org/10.1007/s12530-017-9204-x

Article   CAS   Google Scholar  

Bai T, Tahmasebi P (2021) Attention-based LSTM-FCN for earthquake detection and location. Geophys J Int 228:1568–1576. https://doi.org/10.1093/gji/ggab401

Baiesi M, Paczuski M (2004) Scale-free networks of earthquakes and aftershocks. Phys Rev E Stat Nonlin Soft Matter Phys 69:066106. https://doi.org/10.1103/PhysRevE.69.066106

Article   CAS   PubMed   Google Scholar  

Baker B, Holt MM, Pankow KL et al (2021) Monitoring the 2020 Magna, Utah, earthquake sequence with nodal seismometers and machine learning. Seismol Res Lett 92:787–801. https://doi.org/10.1785/0220200316

Bartlow NM (2020) A long-term view of episodic tremor and slip in Cascadia. Geophys Res Lett 47:e2019GL085303. https://doi.org/10.1029/2019gl085303

Beaucé E, Frank WB, Paul A et al (2019) Systematic detection of clustered seismicity beneath the southwestern alps. J Geophys Res 124:11531–11548. https://doi.org/10.1029/2019jb018110

Bengio Y, Mesnil G, Dauphin Y, Rifai S (2013) Better mixing via deep representations. In: Dasgupta S, McAllester D (eds) Proceedings of the 30th International Conference on Machine Learning. PMLR, Atlanta, Georgia, USA, pp 552–560

Bergen KJ, Beroza GC (2018) Detecting earthquakes over a seismic network using single-station similarity measures. Geophys J Int 213:1984–1998. https://doi.org/10.1093/gji/ggy100

Bergen KJ, Beroza GC (2019) Earthquake fingerprints: extracting waveform features for similarity-based earthquake detection. Pure Appl Geophys 176:1037–1059. https://doi.org/10.1007/s00024-018-1995-6

Bergen K, Yoon C, Beroza GC (2016) Scalable similarity search in seismology: a new approach to large-scale earthquake detection. In: Amsaleg L, Houle ME, Schubert E (eds) Similarity search and applications. Springer International Publishing, Cham, pp 301–308

Chapter   Google Scholar  

Bergen KJ, Johnson PA, de Hoop MV, Beroza GC (2019) Machine learning for data-driven discovery in solid earth geoscience. Science 363:0323. https://doi.org/10.1126/science.aau0323

Bloemheuvel S, van den Hoogen J, Jozinović D et al (2022) Graph neural networks for multivariate time series regression with application to seismic data. International Journal of Data Science and Analytics. https://doi.org/10.1007/s41060-022-00349-6

Bolton DC, Shokouhi P, Rouet-Leduc B et al (2019) Characterizing acoustic signals and searching for precursors during the laboratory seismic cycle using unsupervised machine learning. Seismol Res Lett 90:1088–1098. https://doi.org/10.1785/0220180367

Bolton DC, Shreedharan S, Rivière J, Marone C (2020) Acoustic energy release during the laboratory seismic cycle: insights on laboratory earthquake precursors and prediction. J Geophys Res 125:e2019JB018975. https://doi.org/10.1029/2019JB018975

Borate P, Rivière J, Marone C et al (2023) Using a physics-informed neural network and fault zone acoustic monitoring to predict lab earthquakes. Nat Commun 14:3693. https://doi.org/10.1038/s41467-023-39377-6

Article   CAS   ADS   PubMed   PubMed Central   Google Scholar  

Bozinovski S (2020) Reminder of the first paper on transfer learning in neural networks, 1976. Informatica 44:291–302

Brunsvik B, Morra G, Cambiotti G et al (2021) Three-dimensional paganica fault morphology obtained from hypocenter clustering (L’Aquila 2009 seismic sequence, Central Italy). Tectonophysics 804:228756. https://doi.org/10.1016/j.tecto.2021.228756

Brykov MN, Petryshynets I, Pruncu CI et al (2020) Machine learning modelling and feature engineering in seismology experiment. Sensors 20:4228. https://doi.org/10.3390/s20154228

Article   ADS   PubMed   PubMed Central   Google Scholar  

Bueno Rodriguez A, Benítez C, Zuccarello L et al (2022) Bayesian monitoring of seismo-volcanic dynamics. IEEE Trans Geosci Remote Sens 60:1–14. https://doi.org/10.1109/TGRS.2021.3076012

Campello RJGB, Moulavi D, Sander J (2013) Density-based clustering based on hierarchical density estimates. In: Pei J, Tseng VS, Cao L et al (eds) Advances in knowledge discovery and data mining. Springer, Berlin Heidelberg, pp 160–172

Canário JP, Mello R, Curilem M et al (2020) In-depth comparison of deep artificial neural network architectures on seismic events classification. J Volcanol Geotherm Res 401:106881. https://doi.org/10.1016/j.jvolgeores.2020.106881

Cesca S (2020) Seiscloud, a tool for density-based seismicity clustering and visualization. J Seismol 24:443–457. https://doi.org/10.1007/s10950-020-09921-8

Cesca S, Şen AT, Dahm T (2013) Seismicity monitoring by cluster analysis of moment tensors. Geophys J Int 196:1813–1826. https://doi.org/10.1093/gji/ggt492

Chai C, Maceira M, Santos-Villalobos HJ et al (2020) Using a deep neural network and transfer learning to bridge scales for seismic phase picking. Geophys Res Lett 47:e2020GL088651. https://doi.org/10.1029/2020gl088651

Chai C, Kintner J, Cleveland KM et al (2022) Automatic waveform quality control for surface waves using machine learning. Seismol Res Lett 93:1683–1694. https://doi.org/10.1785/0220210302

Chaipornkaew L, Elston H, Cooke M et al (2022) Predicting off-fault deformation from experimental strike-slip fault images using convolutional neural networks. Geophys Res Lett 49:e2021GL096854. https://doi.org/10.1029/2021gl096854

Chakraborty M, Fenner D, Li W et al (2022) CREIME—a C onvolutional R ecurrent Model for E arthquake I dentification and M agnitude E stimation. J Geophys Res 127:e2022JB024595. https://doi.org/10.1029/2022jb024595

Chamarczuk M, Nishitsuji Y, Malinowski M, Draganov D (2020) Unsupervised learning used in automatic detection and classification of ambient-noise recordings from a large- N Array. Seismol Res Lett 91:370–389. https://doi.org/10.1785/0220190063

Chamberlain CJ, Townend J (2018) Detecting real earthquakes using artificial earthquakes: on the use of synthetic waveforms in matched-filter earthquake detection. Geophys Res Lett 45:11641–11649. https://doi.org/10.1029/2018GL079872

Chamberlain CJ, Hopp CJ, Boese CM et al (2018) EQcorrscan: repeating and near-repeating earthquake detection and analysis in Python. Seismol Res Lett 89:173–181. https://doi.org/10.1785/0220170151

Chen G, Li J (2022) CubeNet: array-based seismic phase picking with deep learning. Seismol Res Lett 93:2554–2569. https://doi.org/10.1785/0220220147

Chen H, Yang H, Zhu G et al (2022a) Deep outer-rise faults in the southern Mariana subduction zone indicated by a machine-learning-based high-resolution earthquake catalog. Geophys Res Lett 49:e2022GL097779. https://doi.org/10.1029/2022gl097779

Chen Y, de Ridder SAL, Rost S et al (2022b) Eikonal tomography with physics-informed neural networks: Rayleigh wave phase velocity in the northeastern margin of the Tibetan plateau. Geophys Res Lett 49:e2022GL099053. https://doi.org/10.1029/2022gl099053

Chen Y, Saad OM, Savvaidis A et al (2022c) 3D microseismic monitoring using machine learning. J Geophys Res 127:e2021JB023842. https://doi.org/10.1029/2021JB023842

Chen KH, Chiu H-Y, Obara K, Liu Y-H (2023) Segmentation characteristics of deep, low-frequency tremors in Shikoku, Japan using machine learning approaches. Earth Planets Space 75:32. https://doi.org/10.1186/s40623-023-01776-w

Cheng Y, Hauksson E, Ben-Zion Y (2023) Refined earthquake focal mechanism catalog for southern California derived with deep learning algorithms. J Geophys Res 128:e2022JB025975. https://doi.org/10.1029/2022jb025975

Chu SX, Beroza GC (2022) Aftershock productivity of intermediate-depth earthquakes in Japan. Geophys J Int 230:448–463. https://doi.org/10.1093/gji/ggac024

Cianetti S, Bruni R, Gaviano S et al (2021) Comparison of deep learning techniques for the investigation of a seismic sequence: an application to the 2019, Mw 45 Mugello (Italy) earthquake. J Geophys Res 126:e2021JB023405. https://doi.org/10.1029/2021jb023405

Civilini F, Weber RC, Jiang Z et al (2021) Detecting moonquakes using convolutional neural networks, a non-local training set, and transfer learning. Geophys J Int 225:2120–2134. https://doi.org/10.1093/gji/ggab083

Article   CAS   ADS   Google Scholar  

Corbi F, Sandri L, Bedford J et al (2019) Machine learning can predict the timing and size of analog earthquakes. Geophys Res Lett 46:1303–1311. https://doi.org/10.1029/2018gl081251

Corbi F, Bedford J, Sandri L et al (2020) Predicting imminence of analog megathrust earthquakes with machine learning: Implications for monitoring subduction zones. Geophys Res Lett 47:e2019GL086615. https://doi.org/10.1029/2019gl086615

Costantino G, Giffard-Roisin S, Marsan D et al (2023) Seismic source characterization from GNSS data using deep learning. J Geophys Res 128:e2022JB024930. https://doi.org/10.1029/2022jb024930

Crowell BW, Bock Y, Liu Z (2016) Single-station automated detection of transient deformation in GPS time series with the relative strength index: a case study of Cascadian slow slip. J Geophys Res 121:9077–9094. https://doi.org/10.1002/2016jb013542

Custódio S, Lima V, Vales D et al (2016) Imaging active faulting in a region of distributed deformation from the joint clustering of focal mechanisms and hypocentres: application to the Azores–western Mediterranean region. Tectonophysics 676:70–89. https://doi.org/10.1016/j.tecto.2016.03.013

Dalai B, Kumar P, Srinu U, Sen MK (2021) De-noising receiver function data using the unsupervised deep learning approach. Geophys J Int 229:737–749. https://doi.org/10.1093/gji/ggab494

Dascher-Cousineau K, Shchur O, Brodsky EE, Günnemann S (2023) Using deep learning for flexible and scalable earthquake forecasting. Geophys Res Lett 50:e2023GL103909. https://doi.org/10.1029/2023gl103909

Datta A, Wu DJ, Zhu W et al (2022) DeepShake: shaking intensity prediction using deep spatiotemporal RNNs for earthquake early warning. Seismol Res Lett 93:1636–1649. https://doi.org/10.1785/0220210141

Derras B, Bard P, Cotton F, Bekkouche A (2012) Adapting the neural network approach to PGA prediction: an example based on the KiK-net data. Bull Seismol Soc Am 102:1446–1461. https://doi.org/10.1785/0120110088

Derras B, Bard PY, Cotton F (2014) Towards fully data driven ground-motion prediction models for Europe. Bull Earthquake Eng 12:495–516. https://doi.org/10.1007/s10518-013-9481-0

DeVries PMR, Thompson TB, Meade BJ (2017) Enabling large-scale viscoelastic calculations via neural network acceleration. Geophys Res Lett 44:2662–2669. https://doi.org/10.1002/2017gl072716

DeVries PMR, Viégas F, Wattenberg M, Meade BJ (2018) Deep learning of aftershock patterns following large earthquakes. Nature 560:632–634. https://doi.org/10.1038/s41586-018-0438-y

Article   CAS   ADS   PubMed   Google Scholar  

Dhanya J, Raghukanth STG (2018) Ground motion prediction model using artificial neural network. Pure Appl Geophys 175:1035–1064. https://doi.org/10.1007/s00024-017-1751-3

Dickey J, Borghetti B, Junek W, Martin R (2020) Beyond correlation: a path-invariant measure for seismogram similarity. Seismol Res Lett 91:356–369. https://doi.org/10.1785/0220190090

Ding W, Li T, Yang X et al (2022) Deep neural networks for creating reliable PmP database with a case study in southern California. J Geophys Res 127:e2021JB023830. https://doi.org/10.1029/2021jb023830

Ding Y, Chen S, Li X et al (2023a) Self-adaptive physics-driven deep learning for seismic wave modeling in complex topography. Eng Appl Artif Intell 123:106425. https://doi.org/10.1016/j.engappai.2023.106425

Ding Y, Chen S, Li X et al (2023b) Physics-constrained neural networks for half-space seismic wave modeling. Comput Geosci 181:105477. https://doi.org/10.1016/j.cageo.2023.105477

Dokht RMH, Kao H, Visser R, Smith B (2019) Seismic event and phase detection using time–frequency representation and convolutional neural networks. Seismol Res Lett 90:481–490. https://doi.org/10.1785/0220180308

Dokht RMH, Kao H, Ghofrani H, Visser R (2022) Combining deep learning and the source-scanning algorithm for improved seismic monitoring. Bull Seismol Soc Am 112:2312–2326. https://doi.org/10.1785/0120220007

Draelos TJ, Ballard S, Young CJ, Brogan R (2015) A new method for producing automated seismic bulletins: probabilistic event detection, association, and location. Bull Seismol Soc Am 105:2453–2467. https://doi.org/10.1785/0120150099

Dupuis M, Schill C, Lee R, Bradley B (2023) A deep-learning-based model for quality assessment of earthquake-induced ground-motion records. Earthq Spectra. https://doi.org/10.1177/87552930231195113

Esfahani RDD, Cotton F, Ohrnberger M, Scherbaum F (2023) TFCGAN: nonstationary ground-motion simulation in the time–frequency domain using conditional generative adversarial network (CGAN) and phase retrieval methods. Bull Seismol Soc Am 113:453–467. https://doi.org/10.1785/0120220068

Esposito AM, Giudicepietro F, D’Auria L et al (2008) Unsupervised neural analysis of very-long-period events at Stromboli volcano using the self-organizing maps. Bull Seismol Soc Am 98:2449–2459. https://doi.org/10.1785/0120070110

Esteghamati MZ, Kottke AR, Rodriguez-Marek A (2022) A data-driven approach to evaluate site amplification of ground-motion models using vector proxies derived from horizontal-to-vertical spectral ratios. Bull Seismol Soc Am 112:3001–3015. https://doi.org/10.1785/0120220106

Ester M, Kriegel H-P, Sander J, Xu X (1996) A density-based algorithm for discovering clusters in large spatial databases with noise. In: International Conference on Knowledge Discovery and Data Mining (KDD). pp 226–231

Faroughi SA, Pawar N, Fernandes C, et al (2022) Physics-guided, physics-informed, and physics-encoded neural networks in scientific computing. arXiv:2211.07377 [cs.LG]

Fauzi A, Mizutani N (2020) Machine learning algorithms for real-time tsunami inundation forecasting: a case study in Nankai region. Pure Appl Geophys 177:1437–1450. https://doi.org/10.1007/s00024-019-02364-4

Feng Q, Han L, Zhao B (2022a) Localizing microseismic events using semi-supervised generative adversarial networks. IEEE Trans Geosci Remote Sens 60:1–8. https://doi.org/10.1109/TGRS.2022.3225415

Feng T, Mohanna S, Meng L (2022b) EdgePhase: a deep learning model for multi-station seismic phase picking. Geochem Geophys Geosyst 23:e2022GC010453. https://doi.org/10.1029/2022GC010453

Festa G, Adinolfi GM, Caruso A et al (2021) Insights into mechanical properties of the 1980 Irpinia fault system from the analysis of a seismic sequence. Geosci J 11:28. https://doi.org/10.3390/geosciences11010028

Florez MA, Caporale M, Buabthong P et al (2022) Data-driven synthesis of broadband earthquake ground motions using artificial intelligence. Bull Seismol Soc Am 112:1979–1996. https://doi.org/10.1785/0120210264

Fornasari SF, Pazzi V, Costa G (2022) A machine-learning approach for the reconstruction of ground-shaking fields in real time. Bull Seismol Soc Am 112:2642–2652. https://doi.org/10.1785/0120220034

Frank WB, Radiguet M, Rousset B et al (2015) Uncovering the geodetic signature of silent slip through repeating earthquakes. Geophys Res Lett 42:2774–2779. https://doi.org/10.1002/2015gl063685

Frohlich C, Pulliam J (1999) Single-station location of seismic events: a review and a plea for more research. Phys Earth Planet Inter 113:277–291. https://doi.org/10.1016/s0031-9201(99)00055-2

Fukushima R, Kano M, Hirahara K (2023) Physics-informed neural networks for fault slip monitoring: simulation, frictional parameter estimation, and prediction on slow slip events in a spring-slider system. J Geophys Res 128:e2023JB027384. https://doi.org/10.1029/2023jb027384

Furumura T, Oishi Y (2023) An early forecast of long-period ground motions of large earthquakes based on deep learning. Geophys Res Lett 50:e2022GL101774. https://doi.org/10.1029/2022gl101774

Furumura M, Ogawa Y, Sakamoto K, Matsu’ura RS, (2023) Automatic digitization of JMA strong-motion seismograms recorded on smoked paper: an attempt using deep learning. Seismol Res Lett. https://doi.org/10.1785/0220230008

Ganter T, Sundermier A, Ballard S (2018) Alternate null hypothesis correlation: a new approach to automatic seismic event detection. Bull Seismol Soc Am 108:3528–3547. https://doi.org/10.1785/0120180074

García JE, Fernández-Prieto LM, Villaseñor A et al (2022) Performance of deep learning pickers in routine network processing applications. Seismol Res Lett 93:2529–2542. https://doi.org/10.1785/0220210323

Garza-Girón R, Brodsky EE, Spica ZJ et al (2023) A specific earthquake processing workflow for studying long-lived, explosive volcanic eruptions with application to the 2008 Okmok Volcano, Alaska, eruption. J Geophys Res 128:e2022JB025882. https://doi.org/10.1029/2022jb025882

Geiger L (1912) Probability method for the determination of earthquake epicentres from the arrival time only. Med Bull St Louis Univ 8:60

Gibbons SJ, Ringdal F (2006) The detection of low magnitude seismic events using array-based waveform correlation. Geophys J Int 165:149–166. https://doi.org/10.1111/j.1365-246X.2006.02865.x

Glasgow M, Schmandt B, Wang R et al (2021) Raton basin induced seismicity is hosted by networks of short basement faults and mimics tectonic earthquake statistics. J Geophys Res 126:e2021JB022839. https://doi.org/10.1029/2021jb022839

Gong J, Fan W, Parnell-Turner R (2022a) Microseismicity indicates atypical small-scale plate rotation at the Quebrada transform fault system, east pacific rise. Geophys Res Lett 49:e2021GL097000. https://doi.org/10.1029/2021gl097000

Gong L-W, Zhang H, Chen S, Chen L-J (2022b) Three-dimensional modeling of the Xichang crust in Sichuan, China by machine learning. NATO Adv Sci Inst Ser E Appl Sci 12:2955. https://doi.org/10.3390/app12062955

Goodfellow IJ, Pouget-Abadie J, Mirza M, et al (2014) Generative adversarial networks. arXiv:1406.2661 [stat.ML]

Granat R, Donnellan A, Heflin M et al (2021) Clustering analysis methods for GNSS observations: a data-driven approach to identifying California’s major faults. Earth Space Sci 8:e2021EA001680. https://doi.org/10.1029/2021EA001680

Grigoli F, Cesca S, Vassallo M, Dahm T (2013) Automated seismic event location by travel-time stacking: an application to mining induced seismicity. Seismol Res Lett 84:666–677. https://doi.org/10.1785/0220120191

Gutierrez LHO, Jiménez CAV, Vásquez LFN (2019) Fast estimation of earthquake arrival azimuth using a single seismological station and machine learning techniques. Earth Sci Res J 23:103–109. https://doi.org/10.15446/esrj.v23n2.70581

Hammer C, Ohrnberger M, Fäh D (2013) Classifying seismic waveforms from scratch: a case study in the alpine environment. Geophys J Int 192:425–439. https://doi.org/10.1093/gji/ggs036

Hara S, Fukahata Y, Iio Y (2019) P-wave first-motion polarity determination of waveform data in western Japan using deep learning. Earth Planets Space 71:1–11. https://doi.org/10.1186/s40623-019-1111-x

He B, Wei M, Watts DR, Shen Y (2020) Detecting slow slip events from seafloor pressure data using machine learning. Geophys Res Lett 47:e2020GL087579. https://doi.org/10.1029/2020gl087579

Heck SL, Young CJ, Brogan R (2022) Comparing traditional and deep learning signal features for event detection in the Utah region. Bull Seismol Soc Am 112:2344–2363. https://doi.org/10.1785/0120210275

Hermkes M, Kuehn NM, Riggelsen C (2014) Simultaneous quantification of epistemic and aleatory uncertainty in GMPEs using Gaussian process regression. Bull Earthquake Eng 12:449–466. https://doi.org/10.1007/s10518-013-9507-7

Hernandez PD, Ramirez JA, Soto MA (2022) Deep-learning-based earthquake detection for fiber-optic distributed acoustic sensing. J Lightwave Technol 40:2639–2650. https://doi.org/10.1109/jlt.2021.3138724

Herrmann M, Kraft T, Tormann T et al (2019) A consistent high-resolution catalog of induced seismicity in Basel based on matched filter detection and tailored post-processing. J Geophys Res 124:8449–8477. https://doi.org/10.1029/2019jb017468

Hicks SP, Goes S, Whittaker AC, Stafford PJ (2021) Multivariate statistical appraisal of regional susceptibility to induced seismicity: application to the Permian basin, SW United States. J Geophys Res [solid Earth] 126:e2021JB022768. https://doi.org/10.1029/2021jb022768

Ho J, Jain A, Abbeel P (2020) Denoising diffusion probabilistic models. arXiv:2006.11239 [cs.LG]

Hoshiba M (2021) Real-time prediction of impending ground shaking: review of wavefield-based (ground-motion-based) method for earthquake early warning. Front Earth Sci 9:722784. https://doi.org/10.3389/feart.2021.722784

Hou X, Zheng Y, Jiang M, Zhang S (2023) SEA-net: sequence attention network for seismic event detection and phase arrival picking. Eng Appl Artif Intell 122:106090. https://doi.org/10.1016/j.engappai.2023.106090

Hu J, Jin C, Zhang H et al (2022) Support vector regression for developing ground-motion models for arias intensity, cumulative absolute velocity, and significant duration for the Kanto region, Japan. Seismol Res Lett 93:1619–1635. https://doi.org/10.1785/0220210259

Hu J, Ding Y, Zhang H et al (2023) A real-time seismic intensity prediction framework based on interpretable ensemble learning. Seismol Res Lett 94:1579–1602. https://doi.org/10.1785/0220220167

Huang X, Alkhalifah T (2022) PINNup: robust neural network wavefield solutions using frequency upscaling and neuron splitting. J Geophys Res 127:e2021JB023703. https://doi.org/10.1029/2021jb023703

Huang X, Alkhalifah T, Song C (2021) A modified physics-informed neural network with positional encoding. In: First International Meeting for Applied Geoscience & Energy Expanded Abstracts. Society of Exploration Geophysicists

Hulbert C, Rouet-Leduc B, Johnson PA et al (2019) Similarity of fast and slow earthquakes illuminated by machine learning. Nat Geosci 12:69–74. https://doi.org/10.1038/s41561-018-0272-8

Ida Y, Ishida M (2022) Analysis of seismic activity using self-organizing map: implications for earthquake prediction. Pure Appl Geophys 179:1–9. https://doi.org/10.1007/s00024-021-02916-7

Ida Y, Fujita E, Hirose T (2022) Classification of volcano-seismic events using waveforms in the method of k-means clustering and dynamic time warping. J Volcanol Geotherm Res 429:107616. https://doi.org/10.1016/j.jvolgeores.2022.107616

Ide S (2021) Empirical low-frequency earthquakes synthesized from tectonic tremor records. J Geophys Res 126:e2021JB022498. https://doi.org/10.1029/2021JB022498

Igel H (2017) Computational seismology: a practical introduction. Oxford University Press, Oxford

Imai R, Kasui N, Iwaki A, Fujiwara H (2021) A sample generation of scenario earthquake shaking maps via a combination of modal decomposition and empirical Copula toward seismic hazard assessment. Bull Seismol Soc Am 111:3341–3355. https://doi.org/10.1785/0120210086

Izzatullah M, Yildirim IE, Waheed UB, Alkhalifah T (2022) Laplace HypoPINN: physics-informed neural network for hypocenter localization and its predictive uncertainty. Mach Learn Sci Technol 3:045001. https://doi.org/10.1088/2632-2153/ac94b3

Jakubowski J, Tajduś A (2014) Predictive regression models of monthly seismic energy emissions induced by longwall mining. Arch Min Sci 59:705–720. https://doi.org/10.2478/amsc-2014-0049

Jasperson H, Bolton DC, Johnson P et al (2021) Attention network forecasts time-to-failure in laboratory shear experiments. J Geophys Res 126:e2021JB022195. https://doi.org/10.1029/2021jb022195

Jebara T (2012) Machine learning: discriminative and generative. Springer, US

Jenkins WF II, Gerstoft P, Bianco MJ, Bromirski PD (2021) Unsupervised deep clustering of seismic data: monitoring the Ross ice shelf, Antarctica. J Geophys Res 126:e2021JB021716. https://doi.org/10.1029/2021jb021716

Ji D, Li C, Zhai C et al (2021) Prediction of ground-motion parameters for the NGA-West2 database using refined second-order deep neural networks. Bull Seismol Soc Am 111:3278–3296. https://doi.org/10.1785/0120200388

Jia L, Chen H, Xing K (2022) Rapid classification of local seismic events using machine learning. J Seismol 26:897–912. https://doi.org/10.1007/s10950-022-10109-5

Jiang C, Zhang P, White MCA et al (2022) A detailed earthquake catalog for Banda Arc-Australian plate collision zone using machine-learning phase picker and an automated workflow. Seism Rec 2:1–10. https://doi.org/10.1785/0320210041

Johnson CW, Ben-Zion Y, Meng H, Vernon F (2020) Identifying different classes of seismic noise signals using unsupervised learning. Geophys Res Lett 47:e2020GL088353. https://doi.org/10.1029/2020gl088353

Johnson PA, Rouet-Leduc B, Pyrak-Nolte LJ et al (2021) Laboratory earthquake forecasting: a machine learning competition. Proc Natl Acad Sci USA 118:e2011362118. https://doi.org/10.1073/pnas.2011362118

Article   CAS   PubMed   PubMed Central   Google Scholar  

Jospin LV, Laga H, Boussaid F et al (2022) Hands-on Bayesian neural networks—A tutorial for deep learning users. IEEE Comput Intell Mag 17:29–48. https://doi.org/10.1109/MCI.2022.3155327

Jozinović D, Lomax A, Štajduhar I, Michelini A (2020) Rapid prediction of earthquake ground shaking intensity using raw waveform data and a convolutional neural network. Geophys J Int 222:1379–1389. https://doi.org/10.1093/gji/ggaa233

Jozinović D, Lomax A, Štajduhar I, Michelini A (2022) Transfer learning: improving neural network based prediction of earthquake ground shaking for an area with insufficient training data. Geophys J Int 229:704–718. https://doi.org/10.1093/gji/ggab488

Kamer Y, Ouillon G, Sornette D (2020) Fault network reconstruction using agglomerative clustering: applications to southern Californian seismicity. Nat Hazards Earth Syst Sci 20:3611–3625. https://doi.org/10.5194/nhess-20-3611-2020

Kamiya M, Igarashi Y, Okada M, Baba T (2022) Numerical experiments on tsunami flow depth prediction for clustered areas using regression and machine learning models. Earth Planets Space 74:127. https://doi.org/10.1186/s40623-022-01680-9

Kaneko R, Nagao H, Ito S-I et al (2021) Convolutional neural network to detect deep low-frequency tremors from seismic waveform images. In: Gupta M, Ramakrishnan G (eds) Lecture notes in computer science. Springer International Publishing, Cham, pp 31–43

Kaneko R, Nagao H, Ito S-I et al (2023) Detection of deep low-frequency tremors from continuous paper records at a station in southwest Japan about 50 years ago based on convolutional neural network. J Geophys Res 128:e2022JB024842. https://doi.org/10.1029/2022jb024842

Karimzadeh S, Matsuoka M, Kuang J, Ge L (2019) Spatial prediction of aftershocks triggered by a major earthquake: a binary machine learning perspective. ISPRS Int J Geoinf 8:462. https://doi.org/10.3390/ijgi8100462

Karniadakis GE, Kevrekidis IG, Lu L et al (2021) Physics-informed machine learning. nature reviews. Physics 3:422–440. https://doi.org/10.1038/s42254-021-00314-5

Kato A, Obara K, Igarashi T et al (2012) Propagation of slow slip leading up to the 2011 M(w) 9.0 Tohoku-Oki earthquake. Science 335:705–708. https://doi.org/10.1126/science.1215141

Kato S (2023) Application of deep learning to automatic seismic wave processing and study of seismic wave reflectors in the deeper part of the Arima-Takatsuki fault zone. Ph D thesis, Kyoto University. https://doi.org/10.14989/doctor.k24308

Kemna KB, Roth MP, Wache RM et al (2022) Small magnitude events highlight the correlation between hydraulic fracturing injection parameters, geological factors, and earthquake occurrence. Geophys Res Lett 49:e2022GL099995. https://doi.org/10.1029/2022gl099995

Kendall K, Tabor D (1971) An utrasonic study of the area of contact between stationary and sliding surfaces. Proc R Soc Lond 323:321–340. https://doi.org/10.1098/rspa.1971.0108

Kerh T, Ting SB (2005) Neural network estimation of ground peak acceleration at stations along Taiwan high-speed rail system. Eng Appl Artif Intell 18:857–866. https://doi.org/10.1016/j.engappai.2005.02.003

Khosravikia F, Clayton P (2021) Machine learning in ground motion prediction. Comput Geosci 148:104700. https://doi.org/10.1016/j.cageo.2021.104700

Kim G, Ku B, Ko H (2021a) Multifeature fusion-based earthquake event classification using transfer learning. IEEE Geosci Remote Sens Lett 18:974–978. https://doi.org/10.1109/LGRS.2020.2993302

Kim S, Yoon B, Lim J-T, Kim M (2021b) Data-driven signal–noise classification for microseismic data using machine learning. Energies 14:1499. https://doi.org/10.3390/en14051499

Kim A, Nakamura Y, Yukutake Y et al (2023) Development of a high-performance seismic phase picker using deep learning in the Hakone volcanic area. Earth Planets Space 75:85. https://doi.org/10.1186/s40623-023-01840-5

Kodera Y, Yamada Y, Hirano K et al (2018) The propagation of local undamped motion (PLUM) method: a simple and robust seismic wavefield estimation approach for earthquake early warning. Bull Seismol Soc Am 108:983–1003. https://doi.org/10.1785/0120170085

Kong Q, Trugman DT, Ross ZE et al (2019) Machine learning in seismology: turning data into insights. Seismol Res Lett 90:3–14. https://doi.org/10.1785/0220180259

Kong Q, Wang R, Walter WR et al (2022) Combining deep learning with physics based features in explosion-earthquake discrimination. Geophys Res Lett 49:e2022GL098645. https://doi.org/10.1029/2022gl098645

Kovachki N, Li Z, Liu B, et al (2021) Neural Operator: Learning Maps Between Function Spaces. arXiv:2108.08481 [cs.LG]

Kriegerowski M, Petersen GM, Vasyura-Bathke H, Ohrnberger M (2019) A deep convolutional neural network for localization of clustered earthquakes based on multistation full waveforms. Seismol Res Lett 90:510–516. https://doi.org/10.1785/0220180320

Kuang W, Yuan C, Zhang J (2021) Real-time determination of earthquake focal mechanism via deep learning. Nat Commun 12:1432. https://doi.org/10.1038/s41467-021-21670-x

Kubo H, Kunugi T, Suzuki W et al (2020) Hybrid predictor for ground-motion intensity with machine learning and conventional ground motion prediction equation. Sci Rep 10:11871. https://doi.org/10.1038/s41598-020-68630-x

Kubo H, Kimura T, Shiomi K (2023) Exploratory data analysis of earthquake moment tensor catalog in japan using non-linear graph-based dimensionality reduction. Pure Appl Geophys 180:2689–2703. https://doi.org/10.1007/s00024-023-03296-w

Kuehn NM, Riggelsen C, Scherbaum F (2011) Modeling the joint probability of earthquake, site, and ground-motion parameters using Bayesian networks. Bull Seismol Soc Am 101:235–249. https://doi.org/10.1785/0120100080

Kumar U, Legendre CP, Zhao L, Chao BF (2022) Dynamic time warping as an alternative to windowed cross correlation in seismological applications. Seismol Res Lett 93:1909–1921. https://doi.org/10.1785/0220210288

Langet N, Maggi A, Michelini A, Brenguier F (2014) Continuous kurtosis-based migration for seismic event detection and location, with application to Piton de la Fournaise volcano, La Reunion. Bull Seismol Soc Am 104:229–246. https://doi.org/10.1785/0120130107

Lapins S, Goitom B, Kendall J-M et al (2021) A little data goes a long way: Automating seismic phase arrival picking at Nabro volcano with transfer learning. J Geophys Res 126:e2021JB021910. https://doi.org/10.1029/2021jb021910

Lara F, Lara-Cueva R, Larco JC et al (2021) A deep learning approach for automatic recognition of seismo-volcanic events at the Cotopaxi volcano. J Volcanol Geotherm Res 409:107142. https://doi.org/10.1016/j.jvolgeores.2020.107142

Larson J, Kramar D, Leonard K (2021) A geostatistical analysis of seismicity in Oklahoma using regression trees and neural networks. Phys Geogr 42:334–350. https://doi.org/10.1080/02723646.2020.1762982

Laurenti L, Tinti E, Galasso F et al (2022) Deep learning for laboratory earthquake prediction and autoregressive forecasting of fault zone stress. Earth Planet Sci Lett 598:117825. https://doi.org/10.1016/j.epsl.2022.117825

Lavrentiadis G, Abrahamson NA, Nicolas KM et al (2023) Overview and introduction to development of non-ergodic earthquake ground-motion models. Bull Earthquake Eng 21:5121–5150. https://doi.org/10.1007/s10518-022-01485-x

Lehmann L, Ohrnberger M, Metz M, Heimann S (2023) Accelerating low-frequency ground motion simulation for finite fault sources using neural networks. Geophys J Int 234:2329–2343. https://doi.org/10.1093/gji/ggad239

Lehmann F, Gatti F, Bertin M, Clouteau D (2024) 3D elastic wave propagation with a factorized Fourier neural operator (F-FNO). Comput Methods Appl Mech Eng 420:116718. https://doi.org/10.1016/j.cma.2023.116718

Article   MathSciNet   ADS   Google Scholar  

Lei X, Satoh T (2007) Indicators of critical point behavior prior to rock failure inferred from pre-failure damage. Tectonophysics 431:97–111. https://doi.org/10.1016/j.tecto.2006.04.023

Li Z (2021) Recent advances in earthquake monitoring II: Emergence of next-generation intelligent systems. Earthquake Sci 34:1–10. https://doi.org/10.29382/eqs-2021-0054

Li Z, Meier M-A, Hauksson E et al (2018) Machine learning seismic wave discrimination: application to earthquake early warning. Geophys Res Lett 45:4773–4779. https://doi.org/10.1029/2018gl077870

Li L, Tan J, Schwarz B et al (2020a) Recent advances and challenges of waveform-based seismic location methods at multiple scales. Rev Geophys 58:e2019RG000667. https://doi.org/10.1029/2019rg000667

Li Y, Ku B, Zhang S et al (2020b) Seismic data augmentation based on conditional generative adversarial networks. Sensors 20:6850. https://doi.org/10.3390/s20236850

Li BQ, Smith JD, Ross ZE (2021) Basal nucleation and the prevalence of ascending swarms in Long Valley caldera. Sci Adv 7:eabi8368. https://doi.org/10.1126/sciadv.abi8368

Li W, Chakraborty M, Fenner D et al (2022a) EPick: attention-based multi-scale UNet for earthquake detection and seismic phase picking. Front Earth Sci 10:953007. https://doi.org/10.3389/feart.2022.953007

Li W, Chakraborty M, Sha Y et al (2022b) A study on small magnitude seismic phase identification using 1D deep residual neural network. Artif Intell Geosci 3:115–122. https://doi.org/10.1016/j.aiig.2022.10.002

Li Z, Zhu L, Officer T et al (2022c) A machine-learning-based method of detecting and picking the first P-wave arrivals of acoustic emission events in laboratory experiments. Geophys J Int 230:1818–1823. https://doi.org/10.1093/gji/ggac148

Liao W, Lee E, Mu D et al (2021) ARRU phase picker: attention recurrent-residual U-Net for picking seismic P- and S-phase arrivals. Seismol Res Lett 92:2410–2428. https://doi.org/10.1785/0220200382

Liao W-Y, Lee E-J, Chen D-Y et al (2022a) RED-PAN: real-time earthquake detection and phase-picking with multitask attention network. IEEE Trans Geosci Remote Sens 60:1–11. https://doi.org/10.1109/TGRS.2022.3205558

Liao W-Y, Lee E-J, Mu D, Chen P (2022b) Toward fully autonomous seismic networks: backprojecting deep learning-based phase time functions for earthquake monitoring on continuous recordings. Seismol Res Lett 93:1880–1894. https://doi.org/10.1785/0220210274

Licciardi A, Bletery Q, Rouet-Leduc B et al (2022) Instantaneous tracking of earthquake growth with elastogravity signals. Nature 606:319–324. https://doi.org/10.1038/s41586-022-04672-7

Lilienkamp H, von Specht S, Weatherill G et al (2022) Ground-motion modeling as an image processing task: introducing a neural network based, fully data-driven, and nonergodic approach. Bull Seismol Soc Am 112:1565–1582. https://doi.org/10.1785/0120220008

Limbeck J, Bisdom K, Lanz F et al (2021) Using machine learning for model benchmarking and forecasting of depletion-induced seismicity in the Groningen gas field. Comput Geosci 25:529–551. https://doi.org/10.1007/s10596-020-10023-0

Article   MathSciNet   Google Scholar  

Lin J-T, Melgar D, Thomas AM, Searcy J (2021) Early warning for great earthquakes from characterization of crustal deformation patterns with deep learning. J Geophys Res 126:e2021JB022703. https://doi.org/10.1029/2021JB022703

Linville L, Pankow K, Draelos T (2019) Deep learning models augment analyst decisions for event discrimination. Geophys Res Lett 46:3643–3651. https://doi.org/10.1029/2018gl081119

Liu M, Zhang M, Zhu W et al (2020) Rapid characterization of the July 2019 Ridgecrest, California, earthquake sequence from raw seismic data using machine-learning phase picker. Geophys Res Lett 47:e2019GL086189. https://doi.org/10.1029/2019gl086189

Liu CM, Rim D, Baraldi R, LeVeque RJ (2021) Comparison of machine learning approaches for tsunami forecasting from sparse observations. Pure Appl Geophys 178:5129–5153. https://doi.org/10.1007/s00024-021-02841-9

Liu H, Li S, Song J (2022) Discrimination between earthquake P waves and microtremors via a generative adversarial network. Bull Seismol Soc Am 112:669–679. https://doi.org/10.1785/0120210231

Lomax A, Michelini A, Jozinović D (2019) An investigation of rapid earthquake characterization using single-station waveforms and a convolutional neural network. Seismol Res Lett 90:517–529. https://doi.org/10.1785/0220180311

Lubbers N, Bolton DC, Mohd-Yusof J et al (2018) Earthquake catalog-based machine learning identification of laboratory fault states and the effects of magnitude of completeness. Geophys Res Lett 45:13269–13276. https://doi.org/10.1029/2018gl079712

Lundberg SM, Lee S-I (2017) A unified approach to interpreting model predictions. In: Advances in Neural Information Processing Systems

Lurka A (2021) Spatio-temporal hierarchical cluster analysis of mining-induced seismicity in coal mines using Ward’s minimum variance method. J Appl Geophys 184:104249. https://doi.org/10.1016/j.jappgeo.2020.104249

Ma X, Chen T (2022) Small seismic events in Oklahoma detected and located by machine learning–based models. Bull Seismol Soc Am 112:2859–2869. https://doi.org/10.1785/0120220029

Ma G, Mei J, Gao K et al (2022) Machine learning bridges microslips and slip avalanches of sheared granular gouges. Earth Planet Sci Lett 579:117366. https://doi.org/10.1016/j.epsl.2022.117366

MacQueen J (1967) Some methods for classification and analysis of multivariate observations. In: Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Statistics. University of California Press, pp 281–298

Magrini F, Jozinović D, Cammarano F et al (2020) Local earthquakes detection: a benchmark dataset of 3-component seismograms built on a global scale. Artif Intell Geosci 1:1–10. https://doi.org/10.1016/j.aiig.2020.04.001

Mai H, Audet P (2022) QuakeLabeler: a fast seismic data set creation and annotation toolbox for AI applications. Seismol Res Lett 93:997–1010. https://doi.org/10.1785/0220210290

Majstorović J, Giffard-Roisin S, Poli P (2021) Designing convolutional neural network pipeline for near-fault earthquake catalog extension using single-station waveforms. J Geophys Res 126:e2020JB021566. https://doi.org/10.1029/2020jb021566

Majstorović J, Giffard-Roisin S, Poli P (2022) Interpreting convolutional neural network decision for earthquake detection with feature map visualization, backward optimization and layer-wise relevance propagation methods. Geophys J Int 232:923–939. https://doi.org/10.1093/gji/ggac369

Makinoshima F, Oishi Y, Yamazaki T et al (2021) Early forecasting of tsunami inundation from tsunami and geodetic observation data with convolutional neural networks. Nat Commun 12:2253. https://doi.org/10.1038/s41467-021-22348-0

Mancini S, Segou M, Werner MJ et al (2022) On the use of high-resolution and deep-learning seismic catalogs for short-term earthquake forecasts: potential benefits and current limitations. J Geophys Res 127:e2022JB025202. https://doi.org/10.1029/2022JB025202

Mao GL, Ferrand TP, Li J et al (2022) Unsupervised machine learning reveals slab hydration variations from deep earthquake distributions beneath the northwest Pacific. Commun Earth Environ 3:56. https://doi.org/10.1038/s43247-022-00377-x

Marano GC, Rosso MM, Aloisio A, Cirrincione G (2023) Generative adversarial networks review in earthquake-related engineering fields. Bull Earthquake Eng. https://doi.org/10.1007/s10518-023-01645-7

Mastella G, Corbi F, Bedford J et al (2022) Forecasting surface velocity fields associated with laboratory seismic cycles using deep learning. Geophys Res Lett 49:e2022GL099632. https://doi.org/10.1029/2022gl099632

Matsumoto Y, Yaoyama T, Lee S et al (2023) Fundamental study on probabilistic generative modeling of earthquake ground motion time histories using generative adversarial networks. Japan Arch Rev 6:e12392. https://doi.org/10.1002/2475-8876.12392

McBrearty IW, Beroza GC (2023) Earthquake phase association with graph neural networks. Bull Seismol Soc Am 113:524–547. https://doi.org/10.1785/0120220182

McBrearty IW, Delorey AA, Johnson PA (2019) Pairwise association of seismic arrivals with convolutional neural networks. Seismol Res Lett 90:503–509. https://doi.org/10.1785/0220180326

McKean SH, Priest JA, Dettmer J, Eaton DW (2019) Quantifying fracture networks inferred from microseismic point clouds by a Gaussian mixture model with physical constraints. Geophys Res Lett 46:11008–11017. https://doi.org/10.1029/2019gl083406

McLaskey GC, Kilgore BD (2013) Foreshocks during the nucleation of stick-slip instability. J Geophys Res 118:2982–2997. https://doi.org/10.1002/jgrb.50232

Meier M-A, Ross ZE, Ramachandran A et al (2019) Reliable real-time seismic signal/noise discrimination with machine learning. J Geophys Res [solid Earth] 124:788–800. https://doi.org/10.1029/2018jb016661

Michelini A, Cianetti S, Gaviano S et al (2021) INSTANCE—the Italian seismic dataset for machine learning. Earth Syst Sci Data 13:5509–5544. https://doi.org/10.5194/essd-13-5509-2021

Mignan A (2014) The debate on the prognostic value of earthquake foreshocks: a meta-analysis. Sci Rep 4:4099. https://doi.org/10.1038/srep04099

Mignan A, Broccardo M (2019) One neuron versus deep learning in aftershock prediction. Nature 574:E1–E3

Mignan A, Broccardo M (2020) Neural network applications in earthquake prediction (1994–2019): meta-analytic and statistical insights on their limitations. Seismol Res Lett 91:2330–2342. https://doi.org/10.1785/0220200021

Mirza M, Osindero S (2014) Conditional generative adversarial nets. arXiv:1411.1784 [cs.LG]

Mitsui Y, Watanabe S (2020) Tectonic division in and around the Izu Peninsula based on the soft clustering of surface displacement rate field. Zisin 73:27–35. https://doi.org/10.4294/zisin.2019-5

Moczo P, Kristek J, Gális M (2014) The finite-difference modelling of earthquake motions: waves and ruptures. Cambridge University Press, Cambridge

Book   Google Scholar  

Mohammadi A, Karimzadeh S, Banimahd SA et al (2023) The potential of region-specific machine-learning-based ground motion models: application to Turkey. Soil Dyn Earthq Eng 172:108008. https://doi.org/10.1016/j.soildyn.2023.108008

Monterrubio-Velasco M, Carrasco-Jimenez JC, Rojas O et al (2022) A statistical approach towards fast estimates of moderate-to-large earthquake focal mechanisms. Front Earth Sci 10:743860. https://doi.org/10.3389/feart.2022.743860

Mosher SG, Audet P (2020) Automatic detection and location of seismic events from time-delay projection mapping and neural network classification. J Geophys Res 125:e2020JB019426. https://doi.org/10.1029/2020jb019426

Mousavi SM, Beroza GC (2020) Bayesian-deep-learning estimation of earthquake location from single-station observations. IEEE Trans Geosci Remote Sens 58:8211–8224. https://doi.org/10.1109/TGRS.2020.2988770

Mousavi SM, Beroza GC (2022) Deep-learning seismology. Science 377:eabm4470. https://doi.org/10.1126/science.abm4470

Mousavi SM, Beroza GC (2023) Machine learning in earthquake seismology. Annu Rev Earth Planet Sci 51:105–129. https://doi.org/10.1146/annurev-earth-071822-100323

Mousavi SM, Sheng Y, Zhu W, Beroza GC (2019a) STanford EArthquake Dataset (STEAD): a global data set of seismic signals for AI. IEEE Access 7:179464–179476. https://doi.org/10.1109/ACCESS.2019.2947848

Mousavi SM, Zhu W, Ellsworth W, Beroza G (2019b) Unsupervised clustering of seismic signals using deep convolutional autoencoders. IEEE Geosci Remote Sens Lett 16:1693–1697. https://doi.org/10.1109/LGRS.2019.2909218

Mousavi SM, Zhu W, Sheng Y, Beroza GC (2019c) CRED: a deep residual network of convolutional and recurrent units for earthquake signal detection. Sci Rep 9:10267. https://doi.org/10.1038/s41598-019-45748-1

Mousavi SM, Ellsworth WL, Zhu W et al (2020) Earthquake transformer-an attentive deep-learning model for simultaneous earthquake detection and phase picking. Nat Commun 11:3952. https://doi.org/10.1038/s41467-020-17591-w

Muir JB, Ross ZE (2023) A deep Gaussian process model for seismicity background rates. Geophys J Int 234:427–438. https://doi.org/10.1093/gji/ggad074

Münchmeyer J, Bindi D, Leser U, Tilmann F (2020) The transformer earthquake alerting model: a new versatile approach to earthquake early warning. Geophys J Int 225:646–656. https://doi.org/10.1093/gji/ggaa609

Münchmeyer J, Bindi D, Leser U, Tilmann F (2021) Earthquake magnitude and location estimation from real time seismic waveforms with a transformer network. Geophys J Int 226:1086–1104. https://doi.org/10.1093/gji/ggab139

Münchmeyer J, Woollam J, Rietbrock A et al (2022) Which picker fits my data? A quantitative evaluation of deep learning based seismic pickers. J Geophys Res 127:e2021JB023499. https://doi.org/10.1029/2021jb023499

Murti MA, Junior R, Ahmed AN, Elshafie A (2022) Earthquake multi-classification detection based velocity and displacement data filtering using machine learning algorithms. Sci Rep 12:21200. https://doi.org/10.1038/s41598-022-25098-1

Nagata K, Nakatani M, Yoshida S (2008) Monitoring frictional strength with acoustic wave transmission. Geophys Res Lett 35:L06310. https://doi.org/10.1029/2007gl033146

Nakamura Y (1988) On the urgent earthquake detection and alarm system (UrEDAS). In: Proc. of the 9th World Conference on Earthquake Engineering. pp 673–678

Nakano M, Sugiyama D (2022) Discriminating seismic events using 1D and 2D CNNs: applications to volcanic and tectonic datasets. Earth Planets Space 74:1–10. https://doi.org/10.1186/s40623-022-01696-1

Nakano M, Sugiyama D, Hori T et al (2019) Discrimination of seismic signals from earthquakes and tectonic tremor by applying a convolutional neural network to running spectral images. Seismol Res Lett 90:530–538. https://doi.org/10.1785/0220180279

Naoi M, Nakatani M, Otsuki K et al (2015) Steady activity of microfractures on geological faults loaded by mining stress. Tectonophysics 649:100–114. https://doi.org/10.1016/j.tecto.2015.02.025

Naoi M, Chen Y, Nishihara K et al (2018) Monitoring hydraulically-induced fractures in the laboratory using acoustic emissions and the fluorescent method. Int J Rock Mech Min Sci 104:53–63. https://doi.org/10.1016/j.ijrmms.2018.02.015

Naoi M, Imakita K, Chen Y et al (2022) Source parameter estimation of acoustic emissions induced by hydraulic fracturing in the laboratory. Geophys J Int 231:408–425. https://doi.org/10.1093/gji/ggac202

Naoi M, Hirano S (2024) Efficient similar waveform search using short binary codes obtained through a deep hashing technique. Geophys J Int. https://doi.org/10.1093/gji/ggae061

Nishimura T, Matsuzawa T, Obara K (2013) Detection of short-term slow slip events along the Nankai Trough, southwest Japan, using GNSS data. J Geophys Res 118:3112–3125. https://doi.org/10.1002/jgrb.50222

Novoselov A, Balazs P, Bokelmann G (2022) SEDENOSS: SEparating and DENOising seismic signals with dual-path recurrent neural network architecture. J Geophys Res [solid Earth] 127:e2021JB023183. https://doi.org/10.1029/2021jb023183

Nustes Andrade J, van der Baan M (2021) Real-time analysis and forecasting of the microseismic cloud size: physics-based models versus machine learning. Geophysics 86:175–186. https://doi.org/10.1190/geo2021-0094.1

Oana A, Ishii T, Miyashita Y, Furukawa K (2022) Construction of ground motion evaluation models using supervised machine learning based on strong motion database. J Jpn Assoc Earthquake Eng 22:6_39-6_56. https://doi.org/10.5610/jaee.22.6_39

Ogata Y (1988) Statistical models for earthquake occurrences and residual analysis for point processes. J Am Stat Assoc 83:9. https://doi.org/10.2307/2288914

Okada Y, Nishimura T, Tabei T et al (2022) Development of a detection method for short-term slow slip events using GNSS data and its application to the Nankai subduction zone. Earth Planets Space 74:1–18. https://doi.org/10.1186/s40623-022-01576-8

Okazaki T, Hachiya H, Iwaki A et al (2021a) Simulation of broad-band ground motions with consistent long-period and short-period components using the Wasserstein interpolation of acceleration envelopes. Geophys J Int 227:333–349. https://doi.org/10.1093/gji/ggab225

Okazaki T, Morikawa N, Fujiwara H, Ueda N (2021b) Monotonic neural network for ground-motion predictions to avoid overfitting to recorded sites. Seismol Res Lett 92:3552–3564. https://doi.org/10.1785/0220210099

Okazaki T, Morikawa N, Iwaki A et al (2021c) Ground-motion prediction model based on neural networks to extract site properties from observational records. Bull Seismol Soc Am 111:1740–1753. https://doi.org/10.1785/0120200339

Okazaki T, Ito T, Hirahara K, Ueda N (2022) Physics-informed deep learning approach for modeling crustal deformation. Nat Commun 13:7092. https://doi.org/10.1038/s41467-022-34922-1

Otake R, Kurima J, Goto H, Sawada S (2020) Deep learning model for spatial interpolation of real-time seismic intensity. Seismol Res Lett 91:3433–3443. https://doi.org/10.1785/0220200006

Ouillon G, Sornette D (2011) Segmentation of fault networks determined from spatial clustering of earthquakes. J Geophys Res 116:B02306. https://doi.org/10.1029/2010jb007752

Ouillon G, Ducorbier C, Sornette D (2008) Automatic reconstruction of fault networks from seismicity catalogs: three-dimensional optimal anisotropic dynamic clustering. J Geophys Res 113:B01306. https://doi.org/10.1029/2007jb005032

Özdemir S, Karslıoğlu MO (2019) Soft clustering of GPS velocities from a homogeneous permanent network in Turkey. J Geod 93:1171–1195. https://doi.org/10.1007/s00190-019-01235-z

Pan SJ, Yang Q (2010) A survey on transfer learning. IEEE Trans Knowl Data Eng 22:1345–1359. https://doi.org/10.1109/TKDE.2009.191

Pan D, Miura H, Kanno T et al (2022) Deep-neural-network-based estimation of site amplification factor from microtremor H/V spectral ratio. Bull Seismol Soc Am 112:1630–1646. https://doi.org/10.1785/0120210300

Paolucci R, Gatti F, Infantino M et al (2018) Broadband ground motions from 3D physics-based numerical simulations using artificial neural networks. Bull Seismol Soc Am 108:1272–1286. https://doi.org/10.1785/0120170293

Park Y, Mousavi SM, Zhu W et al (2020) Machine-learning-based analysis of the Guy-Greenbrier, Arkansas earthquakes: a tale of two sequences. Geophys Res Lett 47:e2020GL087032. https://doi.org/10.1029/2020GL087032

Park Y, Beroza GC, Ellsworth WL (2023) A mitigation strategy for the prediction inconsistency of neural phase pickers. Seismol Res Lett 94:1603–1612. https://doi.org/10.1785/0220230003

Pawley S, Schultz R, Playter T et al (2018) The geological susceptibility of induced earthquakes in the Duvernay Play. Geophys Res Lett 45:1786–1793. https://doi.org/10.1002/2017GL076100

Peng Z, Zhao P (2009) Migration of early aftershocks following the 2004 Parkfield earthquake. Nat Geosci 2:877–881. https://doi.org/10.1038/ngeo697

Peng K, Tang Z, Dong L, Sun D (2021) Machine learning based identification of microseismic signals using characteristic parameters. Sensors 21:6967. https://doi.org/10.3390/s21216967

Perol T, Gharbi M, Denolle M (2018) Convolutional neural network for earthquake detection and location. Sci Adv 4:e1700578. https://doi.org/10.1126/sciadv.1700578

Petersen GM, Niemz P, Cesca S et al (2020) Clusty, the waveform-based network similarity clustering toolbox: concept and application to image complex faulting offshore Zakynthos (Greece). Geophys J Int 224:2044–2059. https://doi.org/10.1093/gji/ggaa568

Picozzi M, Iaccarino AG (2021) Forecasting the preparatory phase of induced earthquakes by recurrent neural network. Forecasting 3:17–36. https://doi.org/10.3390/forecast3010002

Piegari E, Herrmann M, Marzocchi W (2022) 3-D spatial cluster analysis of seismic sequences through density-based algorithms. Geophys J Int 230:2073–2088. https://doi.org/10.1093/gji/ggac160

Pilz M, Cotton F, Kotha SR (2020) Data-driven and machine learning identification of seismic reference stations in Europe. Geophys J Int 222:861–873. https://doi.org/10.1093/gji/ggaa199

Piras D, Spurio Mancini A, Ferreira AMG et al (2022) Towards fast machine-learning-assisted Bayesian posterior inference of microseismic event location and source mechanism. Geophys J Int 232:1219–1235. https://doi.org/10.1093/gji/ggac385

Qin Y, Chen T, Ma X, Chen X (2022a) Forecasting induced seismicity in Oklahoma using machine learning methods. Sci Rep 12:9319. https://doi.org/10.1038/s41598-022-13435-3

Qin Y, Chen X, Chen T, Abercrombie RE (2022b) Influence of fault architecture on induced earthquake sequence evolution revealed by high-resolution focal mechanism solutions. J Geophys Res 127:e2022JB025040. https://doi.org/10.1029/2022jb025040

Raghucharan MC, Somala SN, Erteleva O, Rogozhi E (2021) Seismic attenuation model for data gap regions using recorded and simulated ground motions. Nat Hazards 107:423–446. https://doi.org/10.1007/s11069-021-04589-w

Raissi M, Perdikaris P, Karniadakis GE (2019) Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J Comput Phys 378:686–707. https://doi.org/10.1016/j.jcp.2018.10.045

Rasht-Behesht M, Huber C, Shukla K, Karniadakis GE (2022) Physics-informed neural networks (PINNs) for wave propagation and full waveform inversions. J Geophys Res 127:e2021JB023120. https://doi.org/10.1029/2021jb023120

Rekoske JM, Gabriel A-A, May DA (2023) Instantaneous physics-based ground motion maps using reduced-order modeling. J Geophys Res 128:e2023JB026975. https://doi.org/10.1029/2023jb026975

Ren CX, Peltier A, Ferrazzini V et al (2020) Machine learning reveals the seismic signature of eruptive behavior at Piton de la Fournaise volcano. Geophys Res Lett 47:e2019GL085523. https://doi.org/10.1029/2019GL085523

Retailleau L, Saurel J-M, Zhu W et al (2022) A wrapper to use a machine-learning-based algorithm for earthquake monitoring. Seismol Res Lett 93:1673–1682. https://doi.org/10.1785/0220210279

Retailleau L, Saurel J-M, Laporte M et al (2023) Automatic detection for a comprehensive view of Mayotte seismicity. C R Geosci 354:153–170. https://doi.org/10.5802/crgeos.133

Reynen A, Audet P (2017) Supervised machine learning on a network scale: application to seismic event classification and detection. Geophys J Int 210:1394–1409. https://doi.org/10.1093/gji/ggx238

Ridzwan NSM, Yusoff SHM (2023) Machine learning for earthquake prediction: a review (2017–2021). Earth Sci Inf 16:1133–1149. https://doi.org/10.1007/s12145-023-00991-z

Riel B, Simons M, Agram P, Zhan Z (2014) Detecting transient signals in geodetic time series using sparse estimation techniques. J Geophys Res 119:5140–5160. https://doi.org/10.1002/2014jb011077

Rim D, Baraldi R, Liu CM et al (2022) Tsunami early warning from global navigation satellite system data using convolutional neural networks. Geophys Res Lett 49:e2022GL099511. https://doi.org/10.1029/2022gl099511

Ristea N-C, Radoi A (2022) Complex neural networks for estimating epicentral distance, depth, and magnitude of seismic waves. IEEE Geosci Remote Sens Lett 19:1–5. https://doi.org/10.1109/LGRS.2021.3059422

Rombach R, Blattmann A, Lorenz D, et al (2021) High-resolution image synthesis with latent diffusion models. arXiv:2112.10752 [cs.CV]

Rong K, Yoon CE, Bergen KJ et al (2018) Locality-sensitive hashing for earthquake detection: a case study of scaling data-driven science. Proc VLDB Endow 11:1674–1687

Rosenau M, Corbi F, Dominguez S (2017) Analogue earthquakes and seismic cycles: experimental modelling across timescales. Solid Earth 8:597–635. https://doi.org/10.5194/se-8-597-2017

Ross ZE, Meier M-A, Hauksson E (2018a) P wave arrival picking and first-motion polarity determination with deep learning. J Geophys Res [solid Earth] 123:5120–5129. https://doi.org/10.1029/2017jb015251

Ross ZE, Meier M, Hauksson E, Heaton TH (2018b) Generalized seismic phase detection with deep learning. Bull Seismol Soc Am 108:2894–2901. https://doi.org/10.1785/0120180080

Ross ZE, Trugman DT, Hauksson E, Shearer PM (2019a) Searching for hidden earthquakes in Southern California. Science 364:767–771. https://doi.org/10.1126/science.aaw6888

Ross ZE, Yue Y, Meier M-A et al (2019b) PhaseLink: a deep learning approach to seismic phase association. J Geophys Res [solid Earth] 124:856–869. https://doi.org/10.1029/2018jb016674

Ross ZE, Cochran ES, Trugman DT, Smith JD (2020) 3D fault architecture controls the dynamism of earthquake swarms. Science 368:1357–1361. https://doi.org/10.1126/science.abb0779

Article   MathSciNet   CAS   ADS   PubMed   Google Scholar  

Ross ZE, Zhu W, Azizzadenesheli K (2023) Neural mixture model association of seismic phases. arXiv:2301.02597 [physics.geo-ph]

Roten D, Olsen KB (2021) Estimation of site amplification from geotechnical array data using neural networks. Bull Seismol Soc Am 111:1784–1794. https://doi.org/10.1785/0120200346

Rouet-Leduc B, Hulbert C, Lubbers N et al (2017) Machine learning predicts laboratory earthquakes. Geophys Res Lett 44:9276–9282. https://doi.org/10.1002/2017gl074677

Rouet-Leduc B, Hulbert C, Bolton DC et al (2018a) Estimating fault friction from seismic signals in the laboratory. Geophys Res Lett 45:1321–1329. https://doi.org/10.1002/2017gl076708

Rouet-Leduc B, Hulbert C, Johnson PA (2018b) Continuous chatter of the Cascadia subduction zone revealed by machine learning. Nat Geosci 12:75–79. https://doi.org/10.1038/s41561-018-0274-6

Rouet-Leduc B, Hulbert C, McBrearty IW, Johnson PA (2020) Probing slow earthquakes with deep learning. Geophys Res Lett 47:e2019GL085870. https://doi.org/10.1029/2019GL085870

Rouet-Leduc B, Jolivet R, Dalaison M et al (2021) Autonomous extraction of millimeter-scale deformation in InSAR time series using deep learning. Nat Commun 12:6480. https://doi.org/10.1038/s41467-021-26254-3

Rousset B, Campillo M, Lasserre C et al (2017) A geodetic matched filter search for slow slip with application to the Mexico subduction zone. J Geophys Res 122:10498–10514. https://doi.org/10.1002/2017jb014448

Rundle JB, Donnellan A (2020) Nowcasting earthquakes in southern California with machine learning: bursts, swarms, and aftershocks may be related to levels of regional tectonic stress. Earth Space Sci 7:e2020EA001097. https://doi.org/10.1029/2020ea001097

Rundle JB, Donnellan A, Fox G, Crutchfield JP (2022) Nowcasting earthquakes by visualizing the earthquake cycle with machine learning: a comparison of two methods. Surv Geophys 43:483–501. https://doi.org/10.1007/s10712-021-09655-3

Saad OM, Chen Y (2022) CapsPhase: capsule neural network for seismic phase classification and picking. IEEE Trans Geosci Remote Sens 60:1–11. https://doi.org/10.1109/TGRS.2021.3089929

Saad OM, Huang G, Chen Y et al (2021) SCALODEEP: A highly generalized deep learning framework for real-time earthquake detection. J Geophys Res 126:e2020JB021473. https://doi.org/10.1029/2020jb021473

Saad OM, Chen Y, Trugman D et al (2022) Machine learning for fast and reliable source-location estimation in earthquake early warning. IEEE Geosci Remote Sens Lett 19:1–5. https://doi.org/10.1109/lgrs.2022.3142714

Savage JC (2018) Euler-vector clustering of GPS velocities defines microplate geometry in Southwest Japan. J Geophys Res 123:1954–1968. https://doi.org/10.1002/2017jb014874

Savage JC, Simpson RW (2013a) Clustering of GPS velocities in the Mojave Block, southeastern California. J Geophys Res 118:1747–1759. https://doi.org/10.1029/2012JB009699

Savage JC, Simpson RW (2013b) Clustering of velocities in a GPS network spanning the Sierra Nevada Block, the Northern Walker Lane Belt, and the Central Nevada Seismic Belt, California-Nevada. J Geophys Res [solid Earth] 118:4937–4947. https://doi.org/10.1002/jgrb.50340

Savage JC, Wells RE (2015) Identifying block structure in the Pacific Northwest, USA. J Geophys Res [solid Earth] 120:7905–7916. https://doi.org/10.1002/2015jb012277

Scala A, Adinolfi GM, Picozzi M et al (2022) Monitoring the microseismicity through a dense seismic array and a similarity search detection technique: application to the seismic monitoring of Collalto gas-storage. North Italy Energies 15:3504. https://doi.org/10.3390/en15103504

Schoenball M, Ellsworth WL (2017) A systematic assessment of the spatiotemporal evolution of fault activation through induced seismicity in Oklahoma and southern Kansas. J Geophys Res 122:10,189-10,206. https://doi.org/10.1002/2017jb014850

Scotto di Uccio F, Scala A, Festa G et al (2022) Comparing and integrating artificial intelligence and similarity search detection techniques: application to seismic sequences in Southern Italy. Geophys J Int 233:861–874. https://doi.org/10.1093/gji/ggac487

Seo H, Kim J, Kim B (2022) Machine-learning-based surface ground-motion prediction models for South Korea with low-to-moderate seismicity. Bull Seismol Soc Am 112:1549–1564. https://doi.org/10.1785/0120210244

Seydoux L, Balestriero R, Poli P et al (2020) Clustering earthquake signals and background noises in continuous seismic data with unsupervised deep learning. Nat Commun 11:3972. https://doi.org/10.1038/s41467-020-17841-x

Sharma V, Dhanya J, Gade M, Sivasubramonian J (2022) New generalized ANN-based hybrid broadband response spectra generator using physics-based simulations. Nat Hazards 116:1879–1901. https://doi.org/10.1007/s11069-022-05746-5

Shelly DR, Beroza GC, Ide S (2007) Non-volcanic tremor and low-frequency earthquake swarms. Nature 446:305–307. https://doi.org/10.1038/nature05666

Shelly DR, Hardebeck JL, Ellsworth WL, Hill DP (2016) A new strategy for earthquake focal mechanisms using waveform-correlation-derived relative polarities and cluster analysis: application to the 2014 Long Valley Caldera earthquake swarm. J Geophys Res [solid Earth] 121:8622–8641. https://doi.org/10.1002/2016jb013437

Shen H, Shen Y (2021) Array-based convolutional neural networks for automatic detection and 4D localization of earthquakes in Hawai‘i. Seismol Res Lett 92:2961–2971. https://doi.org/10.1785/0220200419

Shi P, Grigoli F, Lanza F et al (2022) MALMI: an automated earthquake detection and location workflow based on machine learning and waveform migration. Seismol Res Lett 93:2467–2483. https://doi.org/10.1785/0220220071

Shokouhi P, Girkar V, Rivière J et al (2021) Deep learning can predict laboratory quakes from active source seismic data. Geophys Res Lett 48:e2021GL093187. https://doi.org/10.1029/2021gl093187

Shreedharan S, Bolton DC, Rivière J, Marone C (2021) Machine learning predicts the timing and shear stress evolution of lab earthquakes using active seismic monitoring of fault zone processes. J Geophys Res 126:e2020JB021588. https://doi.org/10.1029/2020JB021588

Siemuri A, Selvan K, Kuusniemi H et al (2022) A systematic review of machine learning techniques for GNSS use cases. IEEE Trans Aerosp Electron Syst 58:5043–5077. https://doi.org/10.1109/TAES.2022.3219366

Simpson RW, Thatcher W, Savage JC (2012) Using cluster analysis to organize and explore regional GPS velocities. Geophys Res Lett 39:L18307. https://doi.org/10.1029/2012gl052755

Skoumal RJ, Brudzinski MR, Currie BS (2016) An efficient repeating signal detector to investigate earthquake swarms. J Geophys Res [solid Earth] 121:5880–5897. https://doi.org/10.1002/2016jb012981

Skoumal RJ, Hardebeck JL, Shelly DR (2023a) Using corrected and imputed polarity measurements to improve focal mechanisms in a regional earthquake catalog near the mt Lewis fault zone, California. J Geophys Res 128:e2022JB025660. https://doi.org/10.1029/2022jb025660

Skoumal RJ, Shelly DR, Hardebeck JL (2023b) Using machine learning techniques with incomplete polarity datasets to improve earthquake focal mechanism determination. Seismol Res Lett 94:294–304. https://doi.org/10.1785/0220220103

Smith JD, Ross ZE, Azizzadenesheli K, Muir JB (2021) HypoSVI: hypocentre inversion with stein variational inference and physics informed neural networks. Geophys J Int 228:698–710. https://doi.org/10.1093/gji/ggab309

Song C, Alkhalifah TA (2022) Wavefield reconstruction inversion via physics-informed neural networks. IEEE Trans Geosci Remote Sens 60:1–12. https://doi.org/10.1109/TGRS.2021.3123122

Song C, Wang Y (2022a) Simulating seismic multifrequency wavefields with the Fourier feature physics-informed neural network. Geophys J Int 232:1503–1514. https://doi.org/10.1093/gji/ggac399

Song C, Wang Y (2022b) High-frequency wavefield extrapolation using the Fourier neural operator. J Geophys Eng 19:269–282. https://doi.org/10.1093/jge/gxac016

Song C, Alkhalifah T, Waheed UB (2021a) Solving the frequency-domain acoustic VTI wave equation using physics-informed neural networks. Geophys J Int 225:846–859. https://doi.org/10.1093/gji/ggab010

Song C, Alkhalifah T, Waheed UB (2021b) A versatile framework to solve the Helmholtz equation using physics-informed neural networks. Geophys J Int 228:1750–1762. https://doi.org/10.1093/gji/ggab434

Song J, Zhu J, Wang Y, Li S (2022) On-site alert-level earthquake early warning using machine-learning-based prediction equations. Geophys J Int 231:786–800. https://doi.org/10.1093/gji/ggac220

Soubestre J, Shapiro NM, Seydoux L et al (2018) Network-based detection and classification of seismovolcanic tremors: example from the klyuchevskoy volcanic group in Kamchatka. J Geophys Res [solid Earth] 123:564–582. https://doi.org/10.1002/2017jb014726

Spallarossa D, Kotha SR, Picozzi M et al (2018) On-site earthquake early warning: a partially non-ergodic perspective from the site effects point of view. Geophys J Int 216:919–934. https://doi.org/10.1093/gji/ggy470

Sreenath V, Raghukanth STG (2023) Stochastic ground motion models to NGA-West2 and NGA-Sub databases using Bayesian neural network. Earthq Eng Struct Dyn 52:248–267. https://doi.org/10.1002/eqe.3759

Steinberg A, Vasyura-Bathke H, Gaebler P et al (2021) Estimation of seismic moment tensors using variational inference machine learning. J Geophys Res 126:e2021JB022685. https://doi.org/10.1029/2021jb022685

Steinmann R, Seydoux L, Beaucé É, Campillo M (2022a) Hierarchical Exploration of Continuous Seismograms With Unsupervised Learning. J Geophys Res 127:e2021JB022455. https://doi.org/10.1029/2021JB022455

Steinmann R, Seydoux L, Campillo M (2022b) AI-based unmixing of medium and source signatures from seismograms: ground freezing patterns. Geophys Res Lett 49:e2022GL098854. https://doi.org/10.1029/2022GL098854

Sugiyama D, Tsuboi S, Yukutake Y (2021) Application of deep learning-based neural networks using theoretical seismograms as training data for locating earthquakes in the Hakone volcanic region, Japan. Earth Planets Space 73:1–12. https://doi.org/10.1186/s40623-021-01461-w

Sun H, Ross ZE, Zhu W, Azizzadenesheli K (2023) Phase neural operator for multi-station picking of seismic arrivals. Geophys Res Lett 50:e2023GL106434. https://doi.org/10.1029/2023gl106434

Szafranski D, Duan B (2022) A workflow to integrate numerical simulation, machine learning regression and Bayesian inversion for induced seismicity study: Principles and a case study. Pure Appl Geophys 179:3543–3568. https://doi.org/10.1007/s00024-022-03140-7

Takahashi A, Hashimoto M (2022) Cluster analysis of dense GNSS velocity field reveals characteristics associated with regional tectonics in New Zealand. J Geophys Res 127:e2022JB024793. https://doi.org/10.1029/2022jb024793

Takahashi A, Hashimoto M, Hu J-C et al (2019) Hierarchical cluster analysis of dense GPS data and examination of the nature of the clusters associated with regional tectonics in Taiwan. J Geophys Res [solid Earth] 124:5174–5191. https://doi.org/10.1029/2018jb016995

Takahashi H, Tateiwa K, Yano K, Kano M (2021) A convolutional neural network-based classification of local earthquakes and tectonic tremors in Sanriku-oki, Japan, using S-net data. Earth Planets Space 73:1–10. https://doi.org/10.1186/s40623-021-01524-y

Takanami T, Kitagawa G (1988) A new efficient procedure for the estimation of onset times of seismic waves. J Phys Earth 36:267–290. https://doi.org/10.4294/jpe1952.36.267

Tamaribuchi K, Hirose F, Noda A et al (2021) Noise classification for the unified earthquake catalog using ensemble learning: the enhanced image of seismic activity along the Japan Trench by the S-net seafloor network. Earth Planets Space 73:1–19. https://doi.org/10.1186/s40623-021-01411-6

Tamaribuchi K, Kudo S, Shimojo K, Hirose F (2023) Detection of hidden earthquakes after the 2011 Tohoku earthquake by automatic hypocenter determination combined with machine learning. Earth Planets Space 75:1–19. https://doi.org/10.1186/s40623-023-01915-3

Tamhidi A, Kuehn N, Farid Ghahari S et al (2022) Conditioned simulation of ground-motion time series at uninstrumented sites using Gaussian process regression. Bull Seismol Soc Am 112:331–347. https://doi.org/10.1785/0120210054

Tan YJ, Waldhauser F, Ellsworth WL et al (2021) Machine-learning-based high-resolution earthquake catalog reveals how complex fault structures were activated during the 2016–2017 central Italy sequence. Seismic Rec 1:11–19. https://doi.org/10.1785/0320210001

Tanaka R, Naoi M, Chen Y et al (2021) Preparatory acoustic emission activity of hydraulic fracture in granite with various viscous fluids revealed by deep learning technique. Geophys J Int 226:493–510. https://doi.org/10.1093/gji/ggab096

Tang L, Zhang M, Wen L (2020) Support vector machine classification of seismic events in the Tianshan orogenic belt. J Geophys Res 125:e2019JB018132. https://doi.org/10.1029/2019jb018132

Tezcan J, Cheng Q (2012) Support vector regression for estimating earthquake response spectra. Bull Earthquake Eng 10:1205–1219. https://doi.org/10.1007/s10518-012-9350-2

Thatcher W, Savage JC, Simpson RW (2016) The eastern California shear zone as the northward extension of the southern San Andreas fault. J Geophys Res 121:2904–2914. https://doi.org/10.1002/2015jb012678

Thomas AM, Inbal A, Searcy J et al (2021) Identification of low-frequency earthquakes on the San Andreas fault with deep learning. Geophys Res Lett 48:e2021GL093157. https://doi.org/10.1029/2021gl093157

Tian X, Zhang W, Zhang X et al (2020) Comparison of single-trace and multiple-trace polarity determination for surface microseismic data using deep learning. Seismol Res Lett 91:1794–1803. https://doi.org/10.1785/0220190353

Tibi R, Young C, Gonzales A et al (2017) Rapid and robust cross-correlation-based seismic signal identification using an approximate nearest neighbor method. Bull Seismol Soc Am 107:1954–1968. https://doi.org/10.1785/0120170011

Tibi R, Linville L, Young C, Brogan R (2019) Classification of local seismic events in the Utah region: a comparison of amplitude ratio methods with a spectrogram-based machine learning approach. Bull Seismol Soc Am 109:2532–2544. https://doi.org/10.1785/0120190150

Tibi R, Hammond P, Brogan R et al (2021) Deep learning denoising applied to regional distance seismic data in Utah. Bull Seismol Soc Am 111:775–790. https://doi.org/10.1785/0120200292

Tibi R, Young CJ, Porritt RW (2022) Comparative study of the performance of Seismic waveform denoising methods using local and near-regional data. Bull Seismol Soc Am 113:548–561. https://doi.org/10.1785/0120220105

Titos M, Bueno A, García L et al (2020) Classification of isolated volcano-seismic events based on inductive transfer learning. IEEE Geosci Remote Sens Lett 17:869–873. https://doi.org/10.1109/LGRS.2019.2931063

Tobita M (2016) Combined logarithmic and exponential function model for fitting postseismic GNSS time series after 2011 Tohoku-Oki earthquake. Earth Planets Space 68:1–12. https://doi.org/10.1186/s40623-016-0422-4

Tokuda T, Nagao H (2023) Seismic-phase detection using multiple deep learning models for global and local representations of waveforms. Geophys J Int 235:1163–1182. https://doi.org/10.1093/gji/ggad270

Tous R, Alvarado L, Otero B et al (2020) Deep neural networks for earthquake detection and source region estimation in north-central Venezuela. Bull Seismol Soc Am 110:2519–2529. https://doi.org/10.1785/0120190172

Trampert J, Benzi R, Toschi F (2022) Implications of the statistics of seismicity recorded within the Groningen gas field. Neth J Geosci/geol Mijnbouw 101:E9. https://doi.org/10.1017/njg.2022.8

Trugman DT, Ross ZE (2019) Pervasive foreshock activity across southern California. Geophys Res Lett 46:8772–8781. https://doi.org/10.1029/2019GL083725

Trugman DT, Shearer PM (2018) Strong correlation between stress drop and peak ground acceleration for recent M 1–4 earthquakes in the San Francisco Bay Area. Bull Seismol Soc Am 108:929–945. https://doi.org/10.1785/0120170245

Trugman DT, McBrearty IW, Bolton DC et al (2020) The spatiotemporal evolution of granular microslip precursors to laboratory earthquakes. Geophys Res Lett 47:e2020GL088404. https://doi.org/10.1029/2020gl088404

Uchide T (2020) Focal mechanisms of small earthquakes beneath the Japanese islands based on first-motion polarities picked using deep learning. Geophys J Int 223:1658–1671. https://doi.org/10.1093/gji/ggaa401

Uchide T, Shiina T, Imanishi K (2022) Stress map of japan: detailed nationwide crustal stress field inferred from focal mechanism solutions of numerous microearthquakes. J Geophys Res 127:e2022JB024036. https://doi.org/10.1029/2022jb024036

Unglert K, Jellinek AM (2017) Feasibility study of spectral pattern recognition reveals distinct classes of volcanic tremor. J Volcanol Geotherm Res 336:219–244. https://doi.org/10.1016/j.jvolgeores.2017.03.006

van den Ende MPA, Ampuero J-P (2020) Automated seismic source characterization using deep graph neural networks. Geophys Res Lett 47:e2020GL088690. https://doi.org/10.1029/2020gl088690

Van der Maaten L, Hinton G (2008) Visualizing data using t-SNE. J Mach Learn Res 9:2579–2605

Van Houtte C, Bannister S, Holden C et al (2017) The New Zealand strong motion database. BNZSEE 50:1–20. https://doi.org/10.5459/bnzsee.50.1.1-20

Vinard NA, Drijkoningen GG, Verschuur DJ (2022) Localizing microseismic events on field data using a U-Net-based convolutional neural network trained on synthetic data. Geophysics 87:KS33–KS43. https://doi.org/10.1190/geo2020-0868.1

Waldhauser F, Ellsworth WL (2000) A double-difference earthquake location algorithm: method and application to the northern Hayward fault, California. Bull Seismol Soc Am 90:1353–1368. https://doi.org/10.1785/0120000006

Walter JI, Ogwari P, Thiel A et al (2021) easyQuake: putting machine learning to work for your regional seismic network or local earthquake study. Seismol Res Lett 92:555–563. https://doi.org/10.1785/0220200226

Wamriew D, Charara M, Pissarenko D (2022) Joint event location and velocity model update in real-time for downhole microseismic monitoring: a deep learning approach. Comput Geosci 158:104965. https://doi.org/10.1016/j.cageo.2021.104965

Wang H, Zhang J (2023) A deep learning approach for suppressing noise in livestream earthquake data from a large seismic network. Geophys J Int 233:1546–1559. https://doi.org/10.1093/gji/ggad009

Wang Y, Ouillon G, Woessner J et al (2013) Automatic reconstruction of fault networks from seismicity catalogs including location uncertainty. J Geophys Res 118:5956–5975. https://doi.org/10.1002/2013jb010164

Wang J, Xiao Z, Liu C et al (2019a) Deep learning for picking seismic arrival times. J Geophys Res 124:6612–6624. https://doi.org/10.1029/2019jb017536

Wang K, Ellsworth WL, Beroza GC et al (2019b) Seismology with dark data: Image-based processing of analog records using machine learning for the Rangely earthquake control experiment. Seismol Res Lett 90:553–562. https://doi.org/10.1785/0220180298

Wang R, Schmandt B, Zhang M et al (2020a) Injection-induced earthquakes on complex fault zones of the Raton basin illuminated by machine-learning phase picker and dense nodal array. Geophys Res Lett 47:e2020GL088168. https://doi.org/10.1029/2020gl088168

Wang Z, Zentner I, Zio E (2020b) Accounting for uncertainties of magnitude- and site-related parameters on neural network-computed ground-motion prediction equations. Bull Seismol Soc Am 110:629–646. https://doi.org/10.1785/0120180309

Wang T, Trugman D, Lin Y (2021) SeismoGen: seismic waveform synthesis using GAN with application to seismic data augmentation. J Geophys Res 126:e2020JB020077. https://doi.org/10.1029/2020jb020077

Wang B, Kao H, Dokht RMH et al (2022a) Delineating the controlling factors of hydraulic fracturing-induced seismicity in the Northern Montney Play, Northeastern British Columbia, Canada, with machine learning. Seismol Res Lett 93:2439–2450. https://doi.org/10.1785/0220220075

Wang C-Y, Huang T-C, Wu Y-M (2022b) Using LSTM neural networks for Onsite earthquake early warning. Seismol Res Lett 93:814–826. https://doi.org/10.1785/0220210197

Wang H, Alkhalifah T, Waheed UB, Birnie C (2022c) Data-driven microseismic event localization: a application to the Oklahoma Arkoma basin hydraulic fracturing data. IEEE Trans Geosci Remote Sens 60:1–12. https://doi.org/10.1109/TGRS.2021.3120546

Wang K, Ellsworth W, Beroza GC et al (2022d) DevelNet: Earthquake detection on develocorder films with deep learning: application to the rangely earthquake control experiment. Seismol Res Lett 93:2515–2528. https://doi.org/10.1785/0220220066

Wang A, Li S, Lu J et al (2023a) Prediction of PGA in earthquake early warning using a long short-term memory neural network. Geophys J Int 234:12–24. https://doi.org/10.1093/gji/ggad067

Wang T, Bian Y, Zhang Y, Hou X (2023b) Using artificial intelligence methods to classify different seismic events. Seismol Res Lett 94:1–16. https://doi.org/10.1785/0220220055

Weatherill G, Burton PW (2009) Delineation of shallow seismic source zones using K-means cluster analysis, with application to the Aegean region. Geophys J Int 176:565–588. https://doi.org/10.1111/j.1365-246X.2008.03997.x

Wech AG (2021) Cataloging tectonic tremor energy radiation in the Cascadia subduction zone. J Geophys Res 126:e2021JB022523. https://doi.org/10.1029/2021jb022523

Withers KB, Moschetti MP, Thompson EM (2020) A machine learning approach to developing ground motion models from simulated ground motions. Geophys Res Lett 47:e2019GL086690. https://doi.org/10.1029/2019gl086690

Wong WCJ, Zi JP, Yang HF, Su JR (2021) Spatial-temporal evolution of injection induced earthquakes in weiyuan area by machine-learning phase picker and waveform cross-correlation. Earth Planet Phys 5:485–500. https://doi.org/10.26464/epp2021055

Woollam J, Rietbrock A, Bueno A, De Angelis S (2019) Convolutional neural network for seismic phase classification, performance demonstration over a local seismic network. Seismol Res Lett 90:491–502. https://doi.org/10.1785/0220180312

Woollam J, Münchmeyer J, Tilmann F et al (2022) SeisBench—a toolbox for machine learning in seismology. Seismol Res Lett 93:1695–1709. https://doi.org/10.1785/0220210324

Wozniakowska P, Eaton DW (2020) Machine learning-based analysis of geological susceptibility to induced seismicity in the Montney formation, Canada. Geophys Res Lett 47:e2020GL089651. https://doi.org/10.1029/2020gl089651

Wu Y-M, Kanamori H (2005) Rapid assessment of damage potential of earthquakes in Taiwan from the beginning of P waves. Bull Seismol Soc Am 95:1181–1185. https://doi.org/10.1785/0120040193

Wu Y, Wei J, Pan J, Chen P (2019) Research on microseismic source locations based on deep reinforcement learning. IEEE Access 7:39962–39973. https://doi.org/10.1109/ACCESS.2019.2906066

Xiao Z, Wang J, Liu C et al (2021) Siamese earthquake transformer: a pair-input deep-learning model for earthquake detection and phase picking on a seismic array. J Geophys Res 126:e2020jb021444. https://doi.org/10.1029/2020jb021444

Xu W, Zhu Q, Zhao L (2022) GlitchNet: a glitch detection and removal system for SEIS records based on deep learning. Seismol Res Lett 93:2804–2817. https://doi.org/10.1785/0220210361

Xue X, Freymueller JT (2023) Machine learning for single-station detection of transient deformation in GPS time series with a case study of Cascadia slow slip. J Geophys Res. https://doi.org/10.1029/2022jb024859

Yamaga N, Mitsui Y (2019) Machine learning approach to characterize the postseismic deformation of the 2011 Tohoku-Oki earthquake based on recurrent neural network. Geophys Res Lett 46:11886–11892. https://doi.org/10.1029/2019gl084578

Yáñez-Cuadra V, Moreno M, Ortega-Culaciati F et al (2023) Mosaicking Andean morphostructure and seismic cycle crustal deformation patterns using GNSS velocities and machine learning. Front Earth Sci 11:1096238. https://doi.org/10.3389/feart.2023.1096238

Yang S, Hu J, Zhang H, Liu G (2021a) Simultaneous earthquake detection on multiple stations via a convolutional neural network. Seismol Res Lett 92:246–260. https://doi.org/10.1785/0220200137

Yang Y, Gao AF, Castellanos JC et al (2021b) Seismic wave propagation and inversion with neural operators. The Seismic Record 1:126–134. https://doi.org/10.1785/0320210026

Yang Y, Gao AF, Azizzadenesheli K et al (2023) Rapid seismic waveform modeling and inversion with neural operators. IEEE Trans Geosci Remote Sens 61:1–12. https://doi.org/10.1109/TGRS.2023.3264210

Yano K, Kano M (2022) l 1 trend filtering-based detection of short-term slow slip events: application to a GNSS array in southwest japan. J Geophys Res 127:e2021JB023258. https://doi.org/10.1029/2021jb023258

Yano K, Shiina T, Kurata S et al (2021) Graph-partitioning based convolutional neural network for earthquake detection using a seismic array. J Geophys Res 126:e2020JB020269. https://doi.org/10.1029/2020jb020269

Yin J, Denolle MA, He B (2022a) A multitask encoder–decoder to separate earthquake and ambient noise signal in seismograms. Geophys J Int 231:1806–1822. https://doi.org/10.1093/gji/ggac290

Yin X, Liu F, Cai R et al (2022b) Research on seismic signal analysis based on machine learning. NATO Adv Sci Inst Ser E Appl Sci 12:8389. https://doi.org/10.3390/app12168389

Yoon CE, O’Reilly O, Bergen KJ, Beroza GC (2015) Earthquake detection through computationally efficient similarity search. Sci Adv 1:e1501057. https://doi.org/10.1126/sciadv.1501057

Yoon CE, Huang Y, Ellsworth WL, Beroza GC (2017) Seismicity during the initial stages of the guy-Greenbrier, Arkansas, earthquake sequence. J Geophys Res 122:9253–9274. https://doi.org/10.1002/2017jb014946

Yoon CE, Bergen KJ, Rong K et al (2019a) Unsupervised large-scale search for similar earthquake signals. Bull Seismol Soc Am 109:1451–1468. https://doi.org/10.1785/0120190006

Yoon CE, Yoshimitsu N, Ellsworth WL, Beroza GC (2019b) Foreshocks and mainshock nucleation of the 1999 M w 7.1 Hector mine, California, earthquake. J Geophys Res [solid Earth] 124:1569–1582. https://doi.org/10.1029/2018jb016383

Yu Z, Wang W (2022) FastLink: a machine learning and GPU-based fast phase association method and its application to Yangbi Ms 6.4 aftershock sequences. Geophys J Int 230:673–683. https://doi.org/10.1093/gji/ggac088

Zaliapin I, Gabrielov A, Keilis-Borok V, Wong H (2008) Clustering analysis of seismicity and aftershock identification. Phys Rev Lett 101:018501. https://doi.org/10.1103/PhysRevLett.101.018501

Zhang M, Ellsworth WL, Beroza GC (2019) Rapid earthquake association and location. Seismol Res Lett 90:2276–2284. https://doi.org/10.1785/0220190052

Zhang H, Ma C, Pazzi V et al (2020a) Deep convolutional neural network for microseismic signal detection and classification. Pure Appl Geophys 177:5781–5797. https://doi.org/10.1007/s00024-020-02617-7

Zhang X, Zhang J, Yuan C et al (2020b) Locating induced earthquakes with a network of seismic stations in Oklahoma via a deep learning method. Sci Rep 10:1941. https://doi.org/10.1038/s41598-020-58908-5

Zhang H, Innanen KA, Eaton DW (2021a) Inversion for shear-tensile focal mechanisms using an unsupervised physics-guided neural network. Seismol Res Lett 92:2282–2294. https://doi.org/10.1785/0220200420

Zhang X, Chen H, Zhang W et al (2021b) Generalized neural network trained with a small amount of base samples: application to event detection and phase picking in downhole microseismic monitoring. Geophysics 86:KS95–KS108. https://doi.org/10.1190/geo2020-0955.1

Zhang F, Wang R, Chen Y, Chen Y (2022a) Spatiotemporal variations in earthquake triggering mechanisms during multistage hydraulic fracturing in western Canada. J Geophys Res 127:e2022024744. https://doi.org/10.1029/2022jb024744

Zhang H, Melgar D, Sahakian V et al (2022b) Learning source, path and site effects: CNN-based on-site intensity prediction for earthquake early warning. Geophys J Int 231:2186–2204. https://doi.org/10.1093/gji/ggac325

Zhang M, Liu M, Feng T et al (2022c) LOC-FLOW: an end-to-end machine learning-based high-precision earthquake location workflow. Seismol Res Lett 93:2426–2438. https://doi.org/10.1785/0220220019

Zhang Q, Zhang W, Wu X et al (2022d) Deep learning for efficient microseismic location using source migration-based imaging. J Geophys Res 127:e2021JB022649. https://doi.org/10.1029/2021jb022649

Zhang X, Reichard-Flynn W, Zhang M et al (2022e) Spatiotemporal graph convolutional networks for earthquake source characterization. J Geophys Res 127:e2022JB024401. https://doi.org/10.1029/2022JB024401

Zhang T, Trad D, Innanen K (2023) Learning to solve the elastic wave equation with Fourier neural operators. Geophysics 88:T101–T119. https://doi.org/10.1190/geo2022-0268.1

Zheng J, Harris JM, Li D, Al-Rumaih B (2020) SC-PSNET: aA deep neural network for automatic P- and S-phase detection and arrival-time picker using 1C recordings. Geophysics 85:U87–U98. https://doi.org/10.1190/geo2019-0597.1

Zhou Y, Yue H, Kong Q, Zhou S (2019) Hybrid event detection and phase-picking algorithm using convolutional and recurrent neural networks. Seismol Res Lett 90:1079–1087. https://doi.org/10.1785/0220180319

Zhou PC, Ellsworth WL, Yang HF et al (2021) Machine-learning-facilitated earthquake and anthropogenic source detections near the Weiyuan Shale gas blocks Sichuan China. Earth Planet Phys 5:501–519. https://doi.org/10.26464/epp2021053

Zhu W, Beroza GC (2019) PhaseNet: a deep-neural-network-based seismic arrival-time picking method. Geophys J Int 216:261–273. https://doi.org/10.1093/gji/ggy423

Zhu L, Peng Z, McClellan J et al (2019a) Deep learning for seismic phase detection and picking in the aftershock zone of 2008 M7.9 Wenchuan earthquake. Phys Earth Planet Inter 293:106261. https://doi.org/10.1016/j.pepi.2019.05.004

Zhu W, Mousavi SM, Beroza GC (2019b) Seismic signal denoising and decomposition using deep neural networks. IEEE Trans Geosci Remote Sens 57:9476–9488. https://doi.org/10.1109/TGRS.2019.2926772

Zhu W, Mousavi SM, Beroza GC (2020) Chapter Four—seismic signal augmentation to improve generalization of deep neural networks. In: Moseley B, Krischer L (eds) Advances in geophysics. Elsevier, pp 151–177

Zhu W, Hou AB, Yang R et al (2022a) QuakeFlow: a scalable machine-learning-based earthquake monitoring workflow with cloud computing. Geophys J Int 232:684–693. https://doi.org/10.1093/gji/ggac355

Zhu W, McBrearty IW, Mousavi SM et al (2022b) Earthquake phase association using a Bayesian Gaussian mixture model. J Geophys Res 127:e2021JB023249. https://doi.org/10.1029/2021JB023249

Zhu W, Tai KS, Mousavi SM et al (2022c) An end-to-end earthquake detection method for joint phase picking and association using deep learning. J Geophys Res 127:e2021JB023283. https://doi.org/10.1029/2021jb023283

Zhu C, Cotton F, Kawase H, Bradley B (2023a) Separating broad-band site response from single-station seismograms. Geophys J Int 234:2053–2065. https://doi.org/10.1093/gji/ggad187

Zhu C, Cotton F, Kawase H, Nakano K (2023b) How well can we predict earthquake site response so far? Machine learning vs physics-based modeling. Earthq Spectra 39:478–504. https://doi.org/10.1177/87552930221116399

Zou C, Azizzadenesheli K, Ross ZE, Clayton RW (2023) Deep Neural Helmholtz Operators for 3D elastic wave propagation and inversion. arXiv:2311.09608 [physics.geo-ph]

Download references

Acknowledgements

We thank two anonymous reviewers and the editor, Prof. N. Uchida, for their helpful comments. We would like to thank Dr. Tomohisa Okazaki for his insightful comments.

This study was supported by MEXT Project for Seismology toward Research Innovation with Data of Earthquake (STAR-E) Grant Number JPJ010217 and by the Ministry of Education, Culture, Sports, Science and Technology (MEXT) of Japan, under its Earthquake and Volcano Hazards Observation and Research Program.

Author information

Authors and affiliations.

National Research Institute for Earth Science and Disaster Resilience, 3-1 Tennodai, Tsukuba, 305-0006, Japan

Hisahiko Kubo

Department of Earth and Planetary Sciences, Faculty of Science, Hokkaido University, N10W8, Kita-Ku, Sapporo, 060-0810, Japan

Makoto Naoi

Graduate School of Science, Tohoku University, 6-3, Aramaki Aza-Aoba, Aoba-Ku, Sendai, 980-8578, Japan

Masayuki Kano

You can also search for this author in PubMed   Google Scholar

Contributions

Conceptualization: HK, MN, MK. Formal analysis: HK, MN, MK. Investigation: HK, MN, MK. Data curation: HK, MN, MK. Data: HK, MN, MK. Writing—original draft: HK, MN, MK. Writing—review and editing: HK, MN, MK. Visualization: HK, MN. Funding acquisition: HK, MN, MK. All authors have read and approved the final manuscript.

Corresponding author

Correspondence to Hisahiko Kubo .

Ethics declarations

Ethics approval and consent to participate, consent for publication, competing interests.

The authors declare no competing interests.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Kubo, H., Naoi, M. & Kano, M. Recent advances in earthquake seismology using machine learning. Earth Planets Space 76 , 36 (2024). https://doi.org/10.1186/s40623-024-01982-0

Download citation

Received : 23 July 2023

Accepted : 16 February 2024

Published : 28 February 2024

DOI : https://doi.org/10.1186/s40623-024-01982-0

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Earthquake catalog development
• Seismicity analysis
• Ground-motion prediction
• Geodetic data
• Imbalanced data

Risk Prediction of Earthquakes using Machine Learning

Ieee account.

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

Wearable device could aid dogs in helping to predict earthquakes, but more research needed, critics say

Earthquakes are notoriously hard to predict, but dogs could help bridge the gap.

Can dogs help predict when earthquakes occur?

The company behind the PetPace biometric collar says the health data it tracks may also have the capability of forecasting earthquakes.

A dog's pulse, heart rate variability, temperature, respiration and activity are recorded and sent to the PetPace cloud in real-time, according to Asaf Dagan, chief scientist and cofounder of PetPace, the parent company of Animal Alerts, told ABC News. An AI algorithm then determines the dog’s overall stress level.

"The idea is that if we can track the behavior and the anxiety levels of animals ... then we use AI and machine learning advanced models to correlate that with geophysical data like earthquakes of different magnitudes," Dagan said.

MORE: What scientists discovered after simulating an earthquake on a 10-story wood building

There is currently no conclusive evidence, however, to show that dogs can accurately predict tremors.

Earthquake geologist Wendy Bohon expressed skepticism that a smart collar for dogs could aid in predicting earthquakes.

"I have never seen a convincing study that shows that animals can predict earthquakes, or that animals know that an earthquake is going to happen before it happens," Bohon told ABC News, adding that earthquakes are difficult to predict.

The earliest-ever reference to unusual animal behavior prior to a significant earthquake occurred in 373 BC in Greece, according to the U.S. Geological Survey .

Varying species -- such as rats, weasels, snakes and centipedes -- reportedly left their homes and headed for safety several days before a destructive earthquake struck the region, according to the USGS.

A 2012 paper found that 49% of dogs showed a significant increase in anxiety the day before a 6.8 magnitude earthquake struck the Pacific Northwest in February 2001. The dogs may have picked up on foreshocks, the smaller earthquakes that occur before a larger one, or acoustic waves generated by the movement of the tectonic plates.

"Sometimes animals might feel the arrival of the waves from far away that are too small for us to feel," Bohon said. "And so we don't notice those first small waves arriving ... it may seem like the animals are predicting the earthquake, when in reality they're reacting to the first arrivals of the waves that we're just not tuned into."

MORE: How regional fault lines led to such a destructive earthquake in Turkey, Syria

These primary, or "P" waves, travel at a rate of several miles per second from the epicenter, which could make them more noticeable to animals, according to the USGS .

Very few humans notice the smaller "P" waves, which arrive before the larger "S" waves, the USGS said.

Other historical anecdotal evidence points to fish, birds, reptiles and insects exhibiting strange behavior anywhere from weeks to seconds before an earthquake, the USGS said. Cats, elephants and toads have also been observed to exhibit behavior changes before a seismic event, according to the American Kennel Club .

Evidence in recent decades has largely been anecdotal, according to a 2018 paper published in the Bulletin of the Seismological Society of America that examined 180 previous studies.

Researchers from Germany filmed red wood ants that nest along a fault line in the Neuwied Basin, a seismically active region in Eifel, Germany. They studied the ants’ behavior hours before the earthquake. The scientists found that the ants' nocturnal rest phase and daily activity was suppressed ahead of seismic activity, and their standard daily routine did not resume until the next day.

In a 2020 study , researchers at the Max Planck Institute of Animal Behavior and the University of Konstanz in Germany wrote that they precisely measured an increase in activity in a group of farm animals prior to seismic activity.

Related Stories

A wild orangutan used a medicinal plant to treat a wound, scientists say

• May 2, 11:02 AM

Politicians and dog experts vilify South Dakota governor after she writes about killing her dog

• Apr 29, 6:46 PM

A cluster of earthquakes shakes Taiwan after a strong quake killed 13 earlier this month

• Apr 22, 4:40 PM

The scientists measured the activity of multiple cows, dogs and sheep at a farm near the epicenter of the 6.6 magnitude earthquake that struck near the Norcia Basin in central Italy in October 2016. They found strong reactions among the animals before a magnitude 3.8 or greater earthquake when they were housed together in a stable, according to the study. Prior to the earthquake, the animals were outfitted with biologgers and GPS sensors, highly sensitive instruments that record accelerated movements in any direction.

The farm animals appeared to anticipate tremors anywhere from one to 20 hours ahead, according to the study. They reacted earlier when they were closer to the origin, the researchers said.

MORE: Extreme wave in Marshall Islands highlights dangers of climate change, experts say

Dogs are uniquely suited to behavioral response observation. They can smell odors and hear high-pitched noises undetectable to humans, giving them the ability to sense things outside human perception, according to the American Kennel Club.

"Dogs are particularly useful models of this phenomenon due to their acute senses and their close proximity to humans," Rachell Grant, behavioral ecologist at London South Bank University and lead for the Animal Alert project, said in a statement.

MORE: Cat pulled from rubble 15 days after Italy earthquake

The idea that observational data could be used to predict an oncoming earthquake has been touted in the past. Researchers in 2018 proposed that crowdsourcing and social media might help to predict earthquakes and developed a prototype for social media users to upload posts about abnormal animal behavior.

Social media users claimed that animals behaved strangely in the hours before a 7.8 magnitude earthquake struck Turkey in February 2023, killing more more than 21,000 people. The behavior reported included birds flying erratically and dogs howling loudly, The Washington Post reported.

In 2015, researchers noticed changes in animal behavior three weeks before a 7.0 magnitude quake in an Amazon region of Peru in 2011, a study published in the journal Physics and Chemistry of the Earth found.

PetPace, maker of the biometric dog collar, is conducting a study in Lima, Peru, which experiences 90% of all seismic activity around the world due to its location along the Pacific Ring of Fire. If all of the dogs in the study respond in the same timeframe, the algorithm can likely identify a pattern that an earthquake may be on its way, Dagan said. Dogs in a region with little seismic activity will also be monitored as a control group.

MORE: Dog rescued after spending 9 days trapped under rubble left by Italy earthquake

Scientists want as much data as they can get, and this approach could increase the sample size of animal behavior prior to an earthquake that have already been studied, Bohon said.

"We want all the data and information that we can get, so that we can explore every possibility," she said. "But extraordinary claims require extraordinary evidence."

Related Topics

• Earthquakes

ABC News Live

24/7 coverage of breaking news and live events

Guest editorial for the special issue on artificial intelligence, sensing and big data analytics in earthquake engineering

• Published: 26 April 2024

Cite this article

• Eleni Smyrou 1 ,
• İhsan E. Bal 1 &
• Vasilis Sarhosis 2  

168 Accesses

Explore all metrics

Avoid common mistakes on your manuscript.

1 Introduction

New technologies are advancing at an unprecedentedly accelerating pace over the years. The distance humanity has covered in 2200 years, from the Antikythera mechanism of ancient Greek world, the oldest known analogue computer, to the 4-bit first microprocessor in 1971, is not even comparable to the advancement of technology in the last 50 years. This dazzling journey of technological development has impacted all aspects of modern life, including industry.

Earthquake engineering is one of the disciplines that has embraced new technologies. Earthquake engineers, accustomed to dealing with highly nonlinear and dynamic problems that require complex mathematical and often iterative approaches, are called nowadays to summon dexterity on advanced coding, and masteries on statistics and handling of large amount of data. Artificial Intelligence, Sensing Technologies of all sorts, and Big Data Analytics emerge as essential tools for reducing uncertainty, facilitating engineering process and enhancing knowledge. This Special Issue is a manifestation of the fact that the new technologies can be useful for the most challenging problems of earthquake engineering, opening new prospects in the field.

1.1 Objectives of the special issue

This Special Issue is designed to present a selection of articles on “Artificial Intelligence, Sensing and Big Data Analytics in Earthquake Engineering” to the readers of the Bulletin of Earthquake Engineering. There are two basic objectives of this Special Issue. First, it serves as a platform to present applications of new technologies to practical earthquake engineering problems. Second, it creates awareness for the applicability of such technologies and encourages researchers to make use of these emerging methods more intensively.

When using the new technologies, especially those related to artificial intelligence and its derivatives, it is important to present the results in an explainable way. The physical basis of the addressed problem should be traceable throughout the presented work. This was one important criterion when evaluating the papers included in this Special Issue. Furthermore, a variety of problems have been chosen so that the readers can see a wide range of possible and successful applications of the new technologies and get inspired.

In this Special Issue, it was highlighted that the new technologies either largely improve the existing processes or radically alter the classical approaches used in earthquake engineering shifting the paradigm from “less data– brute computational force” towards “more data– smart algorithms” .

2 Overview of the contributions

An interesting application of artificial neural networks was presented by Kalakonas et al. ( 2022a and 2022b ) where earthquake scenarios for building portfolios were analysed. In an earthquake loss estimation process, three basic components are amalgamated to estimate the risk; hazard, exposure and vulnerability. The paper by Kalakonas et al. ( 2022a ) first addresses the hazard component, by trying to replace the classical multi-parameter regression models used for developing ground motion models with an artificial neural network (ANN) ground motion model. In order to achieve that, they used a compiled database from two subsets of the Pan-European strong motion database (Bindi et al., 2014 ) and the NGA-West2 database (Ancheta et al. 2014 ). Their ANN model uses five input parameters as moment magnitude (Mw), depth, Joyner–Boore distance (Rjb), shear wave velocity in the top 30 m (Vs30), and the faulting type. Their model provides as output the RotD50 horizontal components of common intensity measures at 27 periods from 0.01 to 4.0 s. After establishing the ground motion model, Kalakonas et al. ( 2022b ) presents results of their ANN model for risk estimation for a case study in Balkan region, following the 2019 Durres Earthquake M w 6.4. They concluded that in overall, the ANNs led to damage and economic loss estimates closer to the observations.

Earthquake risk models are used for estimating the effects of earthquakes on a portfolio of buildings geographically distributed in a region of interest. Response of a single building to earthquakes may also be of interest sometimes. If this is the case, structural response reconstruction (SRR) would be one of the possible approaches, where the response of a structure to a specific earthquake motion is not calculated but instead predicted. Abdelmalek‑Lee and Burton ( 2023 ) introduced a dual model that uses kriging combined with the extreme gradient boosting (XGBoost) algorithm. Their aim was to reconstruct the seismic response demands in 207 buildings from 35 real earthquakes. They first used kriging for predicting peak ground accelerations (PGAs) at the location of the buildings, which were not instrumented. Then, they used measured PGAs from instrumented building sites with similar features, such as distance, soil type and magnitude. In the second phase, they used the XGBoost algorithm for reconstructing the maximum peak story drift ratio and peak floor acceleration in the buildings that are not instrumented.

Similar to estimating the building response, prediction of rocking blocks is also an important and challenging theme in earthquake engineering, since it concerns issues such as simple structures, heritage buildings, standing industrial objects or non-structural elements in buildings. Achmet et al. ( 2023 ) proposed a supervised machine learning (ML) model for quickly predicting the seismic response of rocking systems. They compared different supervised ML algorithms, such as the k-Nearest Neighbor (k-NN) and the Support Vector Machine (SVM). They compared the performances of different models by using sine pulses and different sets of natural ground motion records. They presented that the prediction accuracy is practically perfect for sine pulses. They also showed that, as far as prediction for natural records is concerned, accurate results were obtained.

A specific application on model updating at a cultural heritage building was presented in this Special Issue by Salachoris et al. ( 2023 ). They take the civic Clock Tower of Rotella as a case study, a structure damaged during the 2016 Central Italy seismic sequence and use a detailed numerical model of the tower. The paper presents a fully automated finite element model (FEM) updating procedure based on genetic algorithms and global optimization. Their results show that it is possible to automatically create a reference numerical model of the structure in its actual health, which then can be used to assess the dynamic performance of the building.

Marano et al. ( 2023 ) present a thorough review on generative adversarial networks (GANs), specifically for earthquake-related engineering fields. GAN models are particularly useful for generating reliable synthetic data that represent the characteristics of the actual sample set. The amount, variety and quality of data play a tremendous role in the success of prediction models. Their study presents a critical state-of-the-art review of GANs, explaining the most recent research into AI-based GAN synthetic generation of ground motion signals and seismic events. They also offer a comprehensive insight into seismic-related geophysical studies and possible use areas of GANs in that field.

When building a numerical model, calibration of the modelling parameters is sometimes possible if field or experimental data are available. This is a particularly useful approach for calibrating numerical models of heritage structures, since collection of data, especially using destructive methods, is limited or sometimes not even possible. Monchetti et al. ( 2023 ) uses historic masonry towers, where experimental data for model updating were available. They use Bayesian inference, which is frequently employed for addressing parameter uncertainty, observation errors and model inadequacy. Comparison of two different methods, namely Bayesian Model Updating (BMU) and Approximate Bayesian Computation (ABC), results in the conclusion that both methods provide similar results, with the ABC method being more flexible. Their study bridges a gap from field data to accurate numerical modelling, which can be helpful in numerous applications such as digital twinning and automated numerical modelling of structures.

Liu et al. ( 2023 ) used knowledge-enhanced neural networks (KENNs) to tackle the seismic analysis of bridges incorporating simplified component elements. The KENN is used for calibrating parameters of a lump plasticity model in their work. They used a long list of experimental results for identifying the key characteristics of reinforced concrete columns and their seismic response. These key parameters were then employed for supervising KENN model. They demonstrated the accuracy and efficiency of the proposed methodology by applying it to rapid seismic response analysis of typical bridges.

Remote sensing is one of the fastest developing fields in the recent years. It is particularly useful for earthquake-induced structural damage on geographically distributed building stocks and cascading geological hazards. In particular, space-born remote sensing can complement field missions for providing valuable estimates of the spread and degree of damage for better coordinated response and recovery operations.

Giardina et al. ( 2023 ) present application of a novel method on integrating very high-resolution Synthetic Aperture Radar (SAR) data with building survey data for estimating building damages, as well as intensity-based detection of landslides. They presented a successful implementation of the proposed methods for 2021 Haiti Earthquake and Tropical Cyclone Grace.

The new computer technologies, both in the hardware and in the software side, are extremely powerful for image processing. One way of using the image-based methods is for collecting geometrical characteristics of a building using only photogrammetric data and automating the processing of this data for a specific purpose. Pantoja-Rosero et al. ( 2023 ) implement automation of the finite element modelling process using image data and machine learning methods. They present an image-based process where finite element model and mesh are automatically generated, equivalent-frame models of the outer walls of free-standing historical masonry buildings are created. They used and processed RGB (red-green-blue) images of the building in question using structure-from-motion algorithms. They created 3D geometries using convolutional neural networks (CNNs), by segmenting the openings and corners. They then used the generated layers for automatically producing a detailed model. They successfully tested their method on several masonry structures with irregular façades and surfaces.

3 Concluding remarks

Earthquake engineering is an inter-disciplinary engineering branch, the primary purpose of which is protecting human lives against catastrophic earthquakes. Whether designing new buildings and analysing existing structures for future earthquakes or organizing post-earthquake response and recovery, complex, often imperative, problems need to be faced. The challenging nature of these problems stem for the randomness of the earthquakes, as well as from the simple fact that there are earthquakes which can push structures up to and beyond their material limits.

New technologies can play a crucial role in advancing earthquake engineering by providing innovative tools and methods. Advanced simulation and modelling, remote sensing and geospatial analyses, sensors and sensor networks, structural health monitoring and many other technologies, which are not presented in this Special Issue, can assist earthquake engineering in tackling the huge societal responsibility of protecting human lives, and in reducing the likelihood of serious injury or property damage.

The Guest Editors hope that this Special Issue will be inspirational for other researchers who are willing to accept the challenge of improving the earthquake engineering field using the new technologies holistically and more intensively.

Abdelmalek-Lee E, Burton H (2023) A dual Kriging-XGBoost model for reconstructing building seismic responses using strong motion data. Bull Earthq Eng. https://doi.org/10.1007/s10518-023-01624-y

Article   Google Scholar  

Achmet Z, Diamantopoulos S, Fragiadakis M (2023) Rapid seismic response prediction of rocking blocks using machine learning. Bull Earthq Eng. https://doi.org/10.1007/s10518-023-01680-4

Ancheta TD et al (2014) NGA-West2 database. Earthq Spectra 30(3):989–1005. https://doi.org/10.1193/070913EQS197M

Bindi D, Massa M, Luzi L, Ameri G, Pacor F, Puglia R, Augliera P (2014) Pan-european ground-motion prediction equations for the average horizontal component of PGA, PGV, and 5%-damped PSA at spectral periods up to 3.0s using the RESORCE dataset. Bull Earthq Eng 12:391–430. https://doi.org/10.1007/s10518-013-9525-5

Giardina G, Macchiarulo V, Foroughnia F et al (2023) Combining remote sensing techniques and field surveys for post-earthquake reconnaissance missions. Bull Earthq Eng. https://doi.org/10.1007/s10518-023-01716-9

Kalakonas P, Silva V (2022a) Earthquake scenarios for building portfolios using artificial neural networks: part I—ground motion modelling. Bull Earthq Eng. https://doi.org/10.1007/s10518-022-01598-3

Kalakonas P, Silva V (2022b) Earthquake scenarios for building portfolios using artificial neural networks: part II—damage and loss assessment. Bull Earthq Eng. https://doi.org/10.1007/s10518-022-01599-2

Liu Z, Guo A, Zhao C et al (2023) Seismic response of bridges employing knowledge-enhanced neural networks for the lumped plasticity modelling of RC piers. Bull Earthq Eng. https://doi.org/10.1007/s10518-023-01825-5

Marano GC, Rosso MM, Aloisio A et al (2023) Generative adversarial networks review in earthquake-related engineering fields. Bull Earthq Eng. https://doi.org/10.1007/s10518-023-01645-7

Monchetti S, Viscardi C, Betti M et al (2023) Comparison between bayesian updating and approximate bayesian computation for model identification of masonry towers through dynamic data. Bull Earthq Eng. https://doi.org/10.1007/s10518-023-01670-6

Pantoja-Rosero BG, Achanta R, Beyer K (2023) Automated image-based generation of finite element models for masonry buildings. Bull Earthq Eng. https://doi.org/10.1007/s10518-023-01726-7

Salachoris GP, Standoli G, Betti M et al (2023) Evolutionary numerical model for cultural heritage structures via genetic algorithms: a case study in central Italy. Bull Earthq Eng. https://doi.org/10.1007/s10518-023-01615-z

Download references

Acknowledgements

Guest Editors would like to thank the contributing authors who have positively responded to the invitation and provided high-quality research papers for this Special Issue. Guest Editors are also thankful to the reviewers who have meticulously assessed the papers for further improving the quality of the Special Issue. Special thanks are due for Prof. Atilla Ansal, Editor-in-Chief of the journal, and Mrs. Petra van Steenbergen, Springer Executive Editor in Earth Sciences, Geography and Environment, for their encouragement and support.

Author information

Authors and affiliations.

Research Centre Built Environment Noorderruimte, Hanze University of Applied Sciences, Groningen, Netherlands

Eleni Smyrou & İhsan E. Bal

School of Civil Engineering, University of Leeds, LS2 9JT, Leeds, UK

Vasilis Sarhosis

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Eleni Smyrou .

Additional information

Publisher’s note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Reprints and permissions

About this article

Smyrou, E., Bal, İ.E. & Sarhosis, V. Guest editorial for the special issue on artificial intelligence, sensing and big data analytics in earthquake engineering. Bull Earthquake Eng (2024). https://doi.org/10.1007/s10518-024-01920-1

Download citation

Published : 26 April 2024

DOI : https://doi.org/10.1007/s10518-024-01920-1

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Find a journal
• Publish with us
• Track your research