Zefeng Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Zefeng Li has a total of 132 co-authors. Here are the top co-authors by number of shared publications:
Zhigang Peng
9 shared publications- Xin Cui
University of Science and Technology of China
8 shared publications - Jun Zhu
University of Science and Technology of China
7 shared publications - Shang Ma
Univerisity of Science and Technology of China
5 shared publications - Xinming Wu
Laboratory of Seismology and Physics of the Earth’s Interior, School of Earth and Space Sciences, University of Science and Technology of China, Hefei, 230026, China
5 shared publications - Yanlan Hu
Stanford University, Unknown, Stanford, California, United States [US
5 shared publications
Latest Publications
- 10 October 2026
- Published by Elsevier in Earth and Planetary Science Letters
- Vol. 696
Diagnosing machine learning for continental heat flow prediction: Insights from synthetic experiments
- 10 October 2026
- Published by Elsevier in Earth and Planetary Science Letters
- Vol. 696
Diagnosing machine learning for continental heat flow prediction: Insights from synthetic experiments
- 28 July 2026
- Published by Seismological Society of America (SSA) in Seismological Research Letters
Underwater Distributed Acoustic Sensing of Ship-Associated Hydrodynamic Wavefields and Surface Gravity Waves
We characterize ship-associated wavefields and its relation to background surface gravity waves (SGW) on two underwater distributed acoustic sensing (DAS) arrays, one ∼11 km cable at the Foziling reservoir in Anhui, China, and another ∼20 km cable offshore seafloor cable near Yantai in Shandong, China. In the Foziling reservoir, ship-generated signals are separated into mechanical and propeller-related tones, and the 2–10 Hz hydrodynamic noise driven by relative motion between the ship and water. At the Yantai offshore ocean-bottom DAS, higher-frequency ship tones decay rapidly, but the 2–10 Hz hydrodynamic waves remain clear. As ships cross the cable, we observe enhanced energy at ∼0.09 to 0.19 Hz and at ∼0.9 to 1.25 Hz in the Foziling Reservoir, and ∼1.3 to 1.6 Hz at Yantai. Frequency–wavenumber analysis indicates that these enhancements are associated with relative propagation directions between the ship-driven hydrodynamic forcing and background SGW, which is likely due to second-order nonlinear interaction between them. These observations offer insights into dynamic waves caused by waveway transportation and help understand the complex wavefields recorded by underwater DAS.
- 28 July 2026
- Published by Seismological Society of America (SSA) in Seismological Research Letters
Underwater Distributed Acoustic Sensing of Ship-Associated Hydrodynamic Wavefields and Surface Gravity Waves
We characterize ship-associated wavefields and its relation to background surface gravity waves (SGW) on two underwater distributed acoustic sensing (DAS) arrays, one ∼11 km cable at the Foziling reservoir in Anhui, China, and another ∼20 km cable offshore seafloor cable near Yantai in Shandong, China. In the Foziling reservoir, ship-generated signals are separated into mechanical and propeller-related tones, and the 2–10 Hz hydrodynamic noise driven by relative motion between the ship and water. At the Yantai offshore ocean-bottom DAS, higher-frequency ship tones decay rapidly, but the 2–10 Hz hydrodynamic waves remain clear. As ships cross the cable, we observe enhanced energy at ∼0.09 to 0.19 Hz and at ∼0.9 to 1.25 Hz in the Foziling Reservoir, and ∼1.3 to 1.6 Hz at Yantai. Frequency–wavenumber analysis indicates that these enhancements are associated with relative propagation directions between the ship-driven hydrodynamic forcing and background SGW, which is likely due to second-order nonlinear interaction between them. These observations offer insights into dynamic waves caused by waveway transportation and help understand the complex wavefields recorded by underwater DAS.
- Open Access
- 21 April 2026
- Published by American Geophysical Union (AGU) in Geophysical Research Letters
- Vol. 53 (8)
Mainshock‐Induced Stress Changes Modulate Initial Aftershocks on Complex Branching Faults of the 2019 Ridgecrest Earthquake
Large earthquakes can activate complex aftershock fault networks. In such systems, what controls the spatiotemporal evolution of early aftershocks remains a critical yet unresolved problem. Here, using the 2019 M 7.1 Ridgecrest earthquake as an example, we partition the first 10 days of aftershocks onto 15 branching faults activated by the mainshock. For each branching fault, we assess the spatial and temporal influence of mainshock‐induced Coulomb failure stress change (ΔCFS) on aftershock patterns and b values. We find that positive ΔCFS may promote aftershock occurrence across multiple faults during the initial 3 days following the mainshock, but this effect diminishes afterward. Moreover, on most branching faults, higher aftershock b values tend to be associated with areas experiencing negative ΔCFS and reduced differential stress. These findings indicate that mainshock‐induced stress plays an important role in controlling initial aftershock generation and size‐frequency characteristics, providing constraints on aftershock forecasting in complex fault systems.
- 21 April 2026
- Published by American Geophysical Union (AGU) in Geophysical Research Letters
- Vol. 53 (8)
Open AccessMainshock‐Induced Stress Changes Modulate Initial Aftershocks on Complex Branching Faults of the 2019 Ridgecrest Earthquake
Large earthquakes can activate complex aftershock fault networks. In such systems, what controls the spatiotemporal evolution of early aftershocks remains a critical yet unresolved problem. Here, using the 2019 M 7.1 Ridgecrest earthquake as an example, we partition the first 10 days of aftershocks onto 15 branching faults activated by the mainshock. For each branching fault, we assess the spatial and temporal influence of mainshock‐induced Coulomb failure stress change (ΔCFS) on aftershock patterns and b values. We find that positive ΔCFS may promote aftershock occurrence across multiple faults during the initial 3 days following the mainshock, but this effect diminishes afterward. Moreover, on most branching faults, higher aftershock b values tend to be associated with areas experiencing negative ΔCFS and reduced differential stress. These findings indicate that mainshock‐induced stress plays an important role in controlling initial aftershock generation and size‐frequency characteristics, providing constraints on aftershock forecasting in complex fault systems.
- Open Access
- 3 February 2026
- Published by Wiley
Mainshock-induced stress changes modulate initial aftershocks on complex branching faults of the 2019 Ridgecrest earthquake
Large earthquakes can activate complex networks of aftershock faults. In such complex fault systems, what controls the spatiotemporal evolution of early aftershocks remains a critical yet unresolved problem. Here, taking the 2019 M 7.1 Ridgecrest earthquake as an example, we partition the first 10-day aftershocks onto fifteen branching faults activated by the mainshock. For each branching fault, we assess the spatial and temporal influence of mainshock-induced Coulomb failure stress change (ΔCFS) on its aftershock pattern and b value. On most branching faults, higher aftershock b values are observed in areas experiencing negative ΔCFS and reduced differential stress. Moreover, positive ΔCFS promotes aftershock occurrence across multiple faults during the inital 3 days following the mainshock but this promotion effect diminishes afterwards. These findings indicate that mainshock-induced stress plays important role in controling the generation of initial aftershocks and their size-frequency characteristics, providing constraints on aftershock forecasting in complex fault systems.
- 3 February 2026
- Published by Wiley
Open AccessMainshock-induced stress changes modulate initial aftershocks on complex branching faults of the 2019 Ridgecrest earthquake
Large earthquakes can activate complex networks of aftershock faults. In such complex fault systems, what controls the spatiotemporal evolution of early aftershocks remains a critical yet unresolved problem. Here, taking the 2019 M 7.1 Ridgecrest earthquake as an example, we partition the first 10-day aftershocks onto fifteen branching faults activated by the mainshock. For each branching fault, we assess the spatial and temporal influence of mainshock-induced Coulomb failure stress change (ΔCFS) on its aftershock pattern and b value. On most branching faults, higher aftershock b values are observed in areas experiencing negative ΔCFS and reduced differential stress. Moreover, positive ΔCFS promotes aftershock occurrence across multiple faults during the inital 3 days following the mainshock but this promotion effect diminishes afterwards. These findings indicate that mainshock-induced stress plays important role in controling the generation of initial aftershocks and their size-frequency characteristics, providing constraints on aftershock forecasting in complex fault systems.
- 18 November 2025
- Published by Seismological Society of America (SSA) in Seismological Research Letters
- Vol. 97 (3) , 2116-2127
Lithosphere Tectonic Regionalization of China and Surrounding Regions from Unsupervised Learning Analysis of Surface-Wave Dispersion Data
Tectonic regionalization characterizes the primary variation of Earth’s lithosphere and is critical for the quantitative investigation of geodynamic evolution. Although classification of lithospheric blocks based on crustal rock types has been successful, tectonic regionalization with information on deep crustal and mantle structures could provide an alternative perspective due to the addition of mantle structure information. Rayleigh-wave phase velocity dispersion data quantify crustal and upper mantle structures and are thus suitable for deep tectonic regionalization. Here, we used a variational autoencoder (VAE), an unsupervised machine learning method, to classify Rayleigh-wave dispersion maps at 8–120 s in China and its surrounding regions. VAE can compress high-dimensional and nonlinear data to a minimal set of latent variables, which allows a quantitative tectonic regionalization. We found that two latent variables are sufficient to represent the general variations of the dispersion maps. The two latent variables are spatially correlated with topography and the effective elastic thickness of the lithosphere. By generating dispersion curves using even samples from the latent space, we can categorize the continental lithosphere in and around China into four distinct types visually, potentially reflecting different lithospheric activities (e.g., magmatic activities) and composition states related to the physical properties of the lithosphere. We also find that dispersion curves with large reconstruction loss are indicative of regions with relatively poor data resolution. The information extracted from dispersion curves enables us to better understand the integral lithospheric structure from shallow to deep.
- 18 November 2025
- Published by Seismological Society of America (SSA) in Seismological Research Letters
- Vol. 97 (3) , 2116-2127
Lithosphere Tectonic Regionalization of China and Surrounding Regions from Unsupervised Learning Analysis of Surface-Wave Dispersion Data
Tectonic regionalization characterizes the primary variation of Earth’s lithosphere and is critical for the quantitative investigation of geodynamic evolution. Although classification of lithospheric blocks based on crustal rock types has been successful, tectonic regionalization with information on deep crustal and mantle structures could provide an alternative perspective due to the addition of mantle structure information. Rayleigh-wave phase velocity dispersion data quantify crustal and upper mantle structures and are thus suitable for deep tectonic regionalization. Here, we used a variational autoencoder (VAE), an unsupervised machine learning method, to classify Rayleigh-wave dispersion maps at 8–120 s in China and its surrounding regions. VAE can compress high-dimensional and nonlinear data to a minimal set of latent variables, which allows a quantitative tectonic regionalization. We found that two latent variables are sufficient to represent the general variations of the dispersion maps. The two latent variables are spatially correlated with topography and the effective elastic thickness of the lithosphere. By generating dispersion curves using even samples from the latent space, we can categorize the continental lithosphere in and around China into four distinct types visually, potentially reflecting different lithospheric activities (e.g., magmatic activities) and composition states related to the physical properties of the lithosphere. We also find that dispersion curves with large reconstruction loss are indicative of regions with relatively poor data resolution. The information extracted from dispersion curves enables us to better understand the integral lithospheric structure from shallow to deep.
- Open Access
- 31 August 2025
- Published by American Geophysical Union (AGU) in Geophysical Research Letters
- Vol. 52 (17)
Global Search of PKP Precursors With Graph Neural Network: Implications for Scatterers in the Lowermost Mantle
Seismic waves preceding the core phase PKIKP, known as PKP precursors, have been used to map small scatterers near the core‐mantle boundary (CMB). However, their low signal‐to‐noise ratio has required manual identification, hindering systematic and comprehensive analysis. Here, we develop GraphCursor, a model based on Graph Neural Networks (GNNs), to detect PKP precursors across multiple stations. Trained on manually picked precursors, GraphCursor identifies 37,512 precursor waveforms from 5,499 global earthquakes (2000–2024). We locate scatterers by stacking multistation probability functions and estimate their strengths relative to global averages. Our exhaustive search produces unprecedentedly abundant scatterers, including new discoveries of those beneath Antarctica. The distribution of most global scatterers largely correlates with ultra‐low velocity zones (ULVZs). GraphCursor's potential extends to identifying other deep seismic phases and aids in promoting the discovery of new deep Earth structures.
- 31 August 2025
- Published by American Geophysical Union (AGU) in Geophysical Research Letters
- Vol. 52 (17)
Open AccessGlobal Search of PKP Precursors With Graph Neural Network: Implications for Scatterers in the Lowermost Mantle
Seismic waves preceding the core phase PKIKP, known as PKP precursors, have been used to map small scatterers near the core‐mantle boundary (CMB). However, their low signal‐to‐noise ratio has required manual identification, hindering systematic and comprehensive analysis. Here, we develop GraphCursor, a model based on Graph Neural Networks (GNNs), to detect PKP precursors across multiple stations. Trained on manually picked precursors, GraphCursor identifies 37,512 precursor waveforms from 5,499 global earthquakes (2000–2024). We locate scatterers by stacking multistation probability functions and estimate their strengths relative to global averages. Our exhaustive search produces unprecedentedly abundant scatterers, including new discoveries of those beneath Antarctica. The distribution of most global scatterers largely correlates with ultra‐low velocity zones (ULVZs). GraphCursor's potential extends to identifying other deep seismic phases and aids in promoting the discovery of new deep Earth structures.
- 27 August 2025
- Published by Seismological Society of America (SSA) in Seismological Research Letters
- Vol. 97 (1) , 425-438
Unsupervised Deep Clustering of Microseismic Signals from the Dalk Glacier in East Antarctica
In Antarctica, seismic monitoring and detection is significant for understanding long-term glacier activity, glacial melting, and other related scientific issues. We deployed 8 three-component short-period seismometers near the grounding line of the Dalk Glacier ice tongue in Southeast Antarctica, collecting seismic signals from 19 October 2022 to 28 February 2023. We applied unsupervised deep learning, using an autoencoder to encode the features of the input spectrograms into low-dimensional latent vectors, which were then input into a Gaussian mixture model for clustering. We ultimately identified four groups of microseismic signals: Group A wideband events (20–100 Hz) are associated with temperature variation; it is a group of crevasse events. Group B includes low-frequency (<20 Hz) events that may be resonances generated by fluid-filled subglacial fractures, corresponding to the minimum and maximum tidal speeds at the Dalk Glacier coast. Group C events are vibrations caused by wind, and group D events are noise from commuter vehicles. In particular, we discuss the high-energy pulse events included in group A as well as the possible potential sources of group B events. Our research results indicate that deep clustering can effectively identify various types of microseismic signals. The icequakes are closely related to glacier activity driven by environmentally related factors. This may improve further study on the internal structure of glaciers.
- 27 August 2025
- Published by Seismological Society of America (SSA) in Seismological Research Letters
- Vol. 97 (1) , 425-438
Unsupervised Deep Clustering of Microseismic Signals from the Dalk Glacier in East Antarctica
In Antarctica, seismic monitoring and detection is significant for understanding long-term glacier activity, glacial melting, and other related scientific issues. We deployed 8 three-component short-period seismometers near the grounding line of the Dalk Glacier ice tongue in Southeast Antarctica, collecting seismic signals from 19 October 2022 to 28 February 2023. We applied unsupervised deep learning, using an autoencoder to encode the features of the input spectrograms into low-dimensional latent vectors, which were then input into a Gaussian mixture model for clustering. We ultimately identified four groups of microseismic signals: Group A wideband events (20–100 Hz) are associated with temperature variation; it is a group of crevasse events. Group B includes low-frequency (<20 Hz) events that may be resonances generated by fluid-filled subglacial fractures, corresponding to the minimum and maximum tidal speeds at the Dalk Glacier coast. Group C events are vibrations caused by wind, and group D events are noise from commuter vehicles. In particular, we discuss the high-energy pulse events included in group A as well as the possible potential sources of group B events. Our research results indicate that deep clustering can effectively identify various types of microseismic signals. The icequakes are closely related to glacier activity driven by environmentally related factors. This may improve further study on the internal structure of glaciers.
- Open Access
- 11 July 2025
- Published by Oxford University Press (OUP) in Geophysical Journal International
- Vol. 242 (3)
Multistage deep clustering of urban ambient noise for seismic imaging—a case study for train-induced seismic noise
SUMMARY: Passive surface wave method is increasingly being applied to urban subsurface exploration due to its non-invasiveness, low cost and high efficiency. However, its imaging quality is often influenced by limited data acquisition time and the heterogeneous distribution of seismic ambient fields in complex urban environments. To extract coherent surface wave signals for seismic imaging in such challenging setting, we developed a multistage urban ambient noise deep clustering framework based on a convolutional autoencoder and deep embedded clustering algorithm. The initial clustering characterizes the distribution patterns of urban noise sources, which informs a secondary, finer clustering to select noise sources optimized for urban seismic imaging. Real-world experiment on the urban train noise field demonstrates our urban noise cluster framework effectively identifies and elucidates the temporal evolution patterns of moving train sources. Compared to traditional data selection methods, our approach yields superior dispersion measurements and significantly attenuates artifacts from the fundamental mode. Furthermore, by employing mode-specific clustering, we successfully capture the refined first overtone, enhancing the accuracy and depth resolution of seismic imaging. This study presents a new perspective to analysing and selecting complex noise sources, significantly advancing seismic imaging and monitoring in alignment with emerging Artificial Intelligence trends.
- 11 July 2025
- Published by Oxford University Press (OUP) in Geophysical Journal International
- Vol. 242 (3)
Open AccessMultistage deep clustering of urban ambient noise for seismic imaging—a case study for train-induced seismic noise
SUMMARY: Passive surface wave method is increasingly being applied to urban subsurface exploration due to its non-invasiveness, low cost and high efficiency. However, its imaging quality is often influenced by limited data acquisition time and the heterogeneous distribution of seismic ambient fields in complex urban environments. To extract coherent surface wave signals for seismic imaging in such challenging setting, we developed a multistage urban ambient noise deep clustering framework based on a convolutional autoencoder and deep embedded clustering algorithm. The initial clustering characterizes the distribution patterns of urban noise sources, which informs a secondary, finer clustering to select noise sources optimized for urban seismic imaging. Real-world experiment on the urban train noise field demonstrates our urban noise cluster framework effectively identifies and elucidates the temporal evolution patterns of moving train sources. Compared to traditional data selection methods, our approach yields superior dispersion measurements and significantly attenuates artifacts from the fundamental mode. Furthermore, by employing mode-specific clustering, we successfully capture the refined first overtone, enhancing the accuracy and depth resolution of seismic imaging. This study presents a new perspective to analysing and selecting complex noise sources, significantly advancing seismic imaging and monitoring in alignment with emerging Artificial Intelligence trends.
- Open Access
- 4 June 2025
- Published by American Geophysical Union (AGU) in Geophysical Research Letters
- Vol. 52 (11)
Does Foreshock Identification Depend on Seismic Monitoring Capability?
Foreshocks, though well‐documented phenomena preceding many large earthquakes, have limited forecasting utility due to their non‐pervasive occurrence and non‐distinctive characteristics. Using California as an example, we investigate how seismic monitoring capability, particularly the completeness magnitude (), influences the inferred proportion of mainshocks with foreshocks (). We test four foreshock identification methods, namely the fixed‐window, nearest neighbor clustering, empirical statistical (ES) methods and the epidemic‐type aftershock sequence (ETAS) model. The fixed‐window method shows decreasing with higher due to the misclassification of background events as foreshocks. In contrast, clustering and ES methods yield relatively stable across different values. The ETAS model suggests that many foreshocks in California are associated with aseismic driving processes, but the identification of the processes diminishes at high . These results show that improved seismic monitoring capability does not significantly increase but is crucial for distinguishing processes driving foreshocks.
- 4 June 2025
- Published by American Geophysical Union (AGU) in Geophysical Research Letters
- Vol. 52 (11)
Open AccessDoes Foreshock Identification Depend on Seismic Monitoring Capability?
Foreshocks, though well‐documented phenomena preceding many large earthquakes, have limited forecasting utility due to their non‐pervasive occurrence and non‐distinctive characteristics. Using California as an example, we investigate how seismic monitoring capability, particularly the completeness magnitude (), influences the inferred proportion of mainshocks with foreshocks (). We test four foreshock identification methods, namely the fixed‐window, nearest neighbor clustering, empirical statistical (ES) methods and the epidemic‐type aftershock sequence (ETAS) model. The fixed‐window method shows decreasing with higher due to the misclassification of background events as foreshocks. In contrast, clustering and ES methods yield relatively stable across different values. The ETAS model suggests that many foreshocks in California are associated with aseismic driving processes, but the identification of the processes diminishes at high . These results show that improved seismic monitoring capability does not significantly increase but is crucial for distinguishing processes driving foreshocks.
- Open Access
- 8 April 2025
- Published by Oxford University Press (OUP) in Geophysical Journal International
- Vol. 241 (3) , 1853-1867
Multistation seismic location via machine learning: application to Oklahoma and Southern California
SUMMARY: Locating earthquakes plays an important role in the study of seismic activity and geological structures. Traditional methods for locating earthquakes mainly rely on waveform matching and traveltime fitting. With the development of artificial intelligence technology, machine learning methods have often been applied to locate earthquakes. However, current machine learning approaches may face challenges related to physical constraints. In this study, we build a 3-D U-Net network with station distribution constraints to locate earthquakes. To improve the generalizability of the network model, we apply data augmentation techniques, including data shifting, selection, rotation and fusion. The location results are evaluated using a testing data set from Oklahoma, showing an average location error of about 5 km. The origin time can be determined based on the earthquake's location and the waveforms recorded by stations through waveform shifting and stacking. This method does not require the complex processing steps of traditional seismic approaches, allowing for rapid earthquake location. Additionally, we apply the network model to data recorded in Southern California through transfer learning for further application. The results show that this new method is stable and generalized, making it applicable to earthquake location problems associated with arbitrary station networks. Furthermore, we discuss the effects of data augmentation, network architecture and the Gaussian radius of labels on the outcomes. These insights help us better understand machine learning algorithms and improve the application of deep learning in earthquake location.
- 8 April 2025
- Published by Oxford University Press (OUP) in Geophysical Journal International
- Vol. 241 (3) , 1853-1867
Open AccessMultistation seismic location via machine learning: application to Oklahoma and Southern California
SUMMARY: Locating earthquakes plays an important role in the study of seismic activity and geological structures. Traditional methods for locating earthquakes mainly rely on waveform matching and traveltime fitting. With the development of artificial intelligence technology, machine learning methods have often been applied to locate earthquakes. However, current machine learning approaches may face challenges related to physical constraints. In this study, we build a 3-D U-Net network with station distribution constraints to locate earthquakes. To improve the generalizability of the network model, we apply data augmentation techniques, including data shifting, selection, rotation and fusion. The location results are evaluated using a testing data set from Oklahoma, showing an average location error of about 5 km. The origin time can be determined based on the earthquake's location and the waveforms recorded by stations through waveform shifting and stacking. This method does not require the complex processing steps of traditional seismic approaches, allowing for rapid earthquake location. Additionally, we apply the network model to data recorded in Southern California through transfer learning for further application. The results show that this new method is stable and generalized, making it applicable to earthquake location problems associated with arbitrary station networks. Furthermore, we discuss the effects of data augmentation, network architecture and the Gaussian radius of labels on the outcomes. These insights help us better understand machine learning algorithms and improve the application of deep learning in earthquake location.