Machine Learning Applications in Seismology

1. Introduction

The comprehension of earthquakes and natural hazards, including volcanic eruptions and landslides, as well as explosions, through observational data is a pivotal activity within the field of seismology. The rapid advancements in seismogram technology have resulted in the accumulation of extensive seismic datasets, presenting significant opportunities for the exploration of seismicity patterns, the physical processes underlying earthquakes, and the elucidation of earthquake mechanisms [1,2,3,4,5,6,7,8]. The availability of such large-scale seismic data can substantially enhance data-driven research endeavors in seismology [9,10,11,12,13]. These data-rich resources can be employed for a variety of analytical and modeling initiatives, thereby assisting seismologists in gaining insights into earthquake mechanisms, forecasting seismic hazards, and formulating strategies for disaster prevention and mitigation.

Recent progress in seismic data acquisition and processing, particularly through the application of machine learning techniques, has proven beneficial for seismologists in identifying signals or patterns that traditional methodologies may overlook [8,14]. For instance, the automatic detection of seismic events via models such as PhaseNet [15] streamlines the processing of seismic data [16,17,18,19,20]. Additionally, the classification of seismic events utilizing convolutional neural networks (CNN) demonstrates greater efficiency compared to conventional feature-based methods [21,22,23]. Furthermore, machine learning approaches to earthquake prediction and early warning systems offer alternative strategies for mitigating earthquake hazards [4,24,25,26]. In summary, machine learning methodologies significantly enhance the capabilities of seismologists in processing seismic data and uncovering the physical mechanisms associated with earthquakes [27].

Over the past two years, we have compiled 15 articles for this Special Issue titled “Machine Learning Applications in Seismology”. These contributions encompass topics such as seismic inversion, earthquake detection, ground motion simulation, focal mechanism analysis, and earthquake early warning and forecasting systems. The articles underscore the necessity of integrating machine learning into seismological research and provide illustrative examples of its application within the discipline.

2. Summary of the Published Articles

The following overview of the published articles in this Special Issue is organized chronologically by publication date.

Bilal et al. (Contribution 1) introduce an innovative approach to earthquake detection that integrates batch normalization with graph convolutional neural networks (BNGCNN). This study highlights the significance of hyper-parameter optimization in enhancing model performance and demonstrates that the BNGCNN model effectively amalgamates local and global features from seismic data, resulting in improved earthquake detection capabilities. The experimental findings indicate that the BNGCNN surpasses existing models, suggesting its potential utility in real-time earthquake monitoring systems.

Johora et al. (Contribution 2) investigate the use of non-destructive seismic wave velocity measurements to predict geotechnical parameters, such as water content and dry density, through artificial neural networks (ANNs). By incorporating seismic wave velocity data, the ANN models exhibit enhanced predictability compared to traditional multilinear regression models, thereby demonstrating the potential for increased efficiency and accuracy in geotechnical evaluations.

Liu et al. (Contribution 3) utilize convolutional neural networks for the automatic classification of filtered displacement time series derived from strong-motion records, thereby improving efficiency over conventional visual inspection techniques. By employing transfer learning with models such as VGG19 and ResNet50, this research work achieves enhanced accuracy in determining high-pass cut-off frequencies, attaining a maximum coefficient of determination (R²) of 0.82 with minimal prediction errors.

Merdasse et al. (Contribution 4) apply time series analysis to forecast earthquake frequency and magnitude in northeastern Algeria, utilizing both parametric (autoregressive integrated moving average, ARIMA) and non-parametric (singular spectrum analysis, SSA) methodologies. Analyzing data from 1910 to 2019, the findings reveal that the SSA model outperforms the ARIMA model. Their forecasts indicate that between 2020 and 2030, the annual maximum magnitude will range from Mw 4.8 to Mw 5.1, with an expectation of four to six earthquakes of at least Mw 4.0 occurring annually.

Li et al. (Contribution 5) propose FocMech-Flow, an automated workflow designed for determining P-wave first-motion polarity and focal mechanism inversion, applied to the 2021 Yangbi earthquake sequence. Utilizing the deep learning model DiTingMotion, the method achieves an accuracy of 98.49% in polarity detection and provides 112 focal mechanism solutions, thereby enhancing the understanding of fault structures and regional stress fields in small to moderate earthquakes.

Agayan et al. (Contribution 6) present advancements in the FCAZ (fuzzy clustering and zoning) method for identifying earthquake-prone regions through enhancements to its mathematical algorithms and foundational principles. This study focuses on improving the precision and reliability of high-seismicity area identification, exemplified by a case study in California. By refining the FCAZ algorithm, closely linked small zones are consolidated into larger, high-seismicity areas, thereby enhancing the efficacy of earthquake hazard assessments.

Wang et al. (Contribution 7) explore the potential of machine learning techniques, specifically random forest and long short-term memory (LSTM) neural networks, to predict large earthquakes utilizing seismic catalog data from the Sichuan–Yunnan region. The research addresses two critical questions: the likelihood of a significant earthquake occurring within a year and the anticipated maximum magnitude. The results indicate that the random forest method excels in classifying large earthquake occurrences, while LSTM provides reasonable magnitude estimations. The findings suggest that small earthquakes contain valuable predictive information, underscoring the promise of machine learning in improving earthquake prediction accuracy.

Hu Junjun et al. (Contribution 8) propose a novel software application employing machine learning to simulate ground motion by accurately matching amplitude, spectrum, and duration characteristics specific to a region. By utilizing principal component analysis and predictive equations, this software generates simulated ground motions that closely align with the desired attributes, offering a more reliable input for structural design and assessment.

Zhu et al. (Contribution 9) investigate anomalies in outgoing longwave radiation (OLR) data preceding the Yangbi Ms6.4 and Luding Ms6.8 earthquakes using the bidirectional long short-term memory (BILSTM) model. This study predicts OLR values prior to the earthquakes, employing confidence intervals for anomaly detection. The authors suggest that their method effectively captures seismic anomalies and may indicate a correlation between OLR anomalies and earthquake occurrences, advocating for further research involving additional earthquake cases to enhance predictive capabilities.

Agathos et al. (Contribution 10) discuss the application of a specialized deep neural network to identify earthquakes in environments characterized by significant background noise from vehicular activity. To address this challenge, this study proposes utilizing a deep neural network trained on both earthquake and vehicular signals to detect earthquakes within low-cost sensor data contaminated by noise, demonstrating superior effectiveness and efficiency compared to traditional models. This article emphasizes the critical role of earthquake monitoring in disaster management, public safety, and scientific inquiry.

Tang et al. (Contribution 11) present a novel seismic inversion method that employs a multi-scale super-asymmetric network (Cycle-JNet) to enhance the resolution and accuracy of seismic data interpretation. By integrating wavelet analysis with deep learning techniques, the Cycle-JNet model effectively captures multi-scale data characteristics. The model exhibits superior performance in identifying thin sandstone layers compared to conventional approaches, achieving a prediction accuracy of 81.2%. The findings indicate that the Cycle-JNet network serves as a reliable tool for seismic inversion, significantly improving the identification of geological features in complex data environments.

Li et al. (Contribution 12) develop a high-resolution aftershock catalog for the 2014 Ms 6.5 Ludian earthquake in China utilizing deep learning methodologies, specifically the deep learning phase-picking CERP model and seismic-phase association PALM technology. A novel training strategy that combines traditional algorithms with artificial intelligence enhances seismic phase detection and event localization, resulting in the identification of 3286 aftershock events with improved accuracy. This study underscores the effectiveness of the retraining strategy in enhancing the generalization of AI models for seismic analysis in specific tectonic contexts.

Li et al. (Contribution 13) present a deep learning approach for microseismic velocity inversion, employing a Unet model in conjunction with data augmentation and hybrid training strategies. This methodology effectively addresses challenges associated with low signal-to-noise ratios in real microseismic data, enhancing inversion accuracy by integrating synthetic and augmented datasets. The results demonstrate the model’s robustness and potential for improved subsurface velocity predictions.

Zhao et al. (Contribution 14) introduce a novel approach for enhancing passive seismic source reconstruction through the use of convolutional neural networks (CNN). The authors tackle the challenges posed by randomly distributed and sparse seismic sources, which often result in artifacts and coherent noise in virtual shot records. By incorporating an adaptive attention mechanism into the CNN architecture, the proposed method effectively suppresses noise and restores valid waveform features. The results demonstrate improved signal clarity and continuity, highlighting the method’s applicability in passive seismic exploration, particularly in scenarios characterized by uneven source distributions and limited active sources.

Lu et al. (Contribution 15) investigate the application of machine learning for detecting earthquake precursors through the analysis of seismic multi-parameter data across twelve tectonic regions in western China. This study employs a sliding extreme value relevancy method to analyze various seismic parameters, including the b value, earthquake frequency, and intensity factors. Their findings indicate that significant anomalies frequently precede target earthquakes, with high anomaly rates correlating with earthquake occurrences. The results emphasize the effectiveness of a comprehensive multi-parameter approach in enhancing earthquake prediction accuracy and spatial risk assessment, providing valuable insights for future evaluations of seismic hazards in the investigated regions.

3. Conclusions

The adoption of artificial intelligence in scientific research has gained traction across various disciplines, achieving notable success. The application of machine learning within seismology has also garnered increasing attention. This collection of articles in this Special Issue contribute to the field of seismology by presenting views and research examples that illustrate the integration of machine learning into seismology. The topics covered in these articles include seismic inversion, earthquake detection, focal mechanism analysis, ground motion simulation, earthquake early warning systems, and earthquake forecasting, utilizing a diverse array of machine learning methods.

Moreover, these articles highlight interdisciplinary research that bridges seismology and machine learning, offering innovative solutions to challenges associated with seismic data and advancements in model interpretability. While limitations exist, there is a strong expectation for future work to focus on enhancing model accuracy and generalizability, the development of real-time applications in seismology and the exploration of the physical mechanisms underlying earthquakes through machine learning methodologies.

In conclusion, we extend our heartfelt appreciation to the contributors, reviewers, and the editorial team for their dedicated efforts in bringing this Special Issue to fruition. Their contributions have significantly enriched the discourse on the application of machine learning in seismology, paving the way for future advancements in the field.