An official website of the United States government
Earthquakes in stable salt domes are few, with a notable increase in the rate of seismicity prior to catastrophic events, such as the collapse of salt caverns used to store hydrocarbons. Cavern collapse, subsequent gas leakage, and the formation of sinkholes pose a significant hazard for local communities, given that they can disrupt normal societal functions, have various socio-economic impacts, and may result in the evacuation of residents. In Louisiana, one such event was the Bayou Corne collapse in 2012. Following reports of unusual ground tremors, we began monitoring seismicity at the Sorrento salt dome in February 2020. The goal of this study is to improve our understanding of the subsurface processes and their impact on the mechanical integrity of salt domes; we do this by examining the spatio-temporal evolution of the seismicity. We deployed an ~5 km x 4 km nodal array of 12-17 stations, with interstation distances of 0.2 km to 1.9 km, across the dome and recorded eight months of data that were sampled at 500 Hz. Sorrento dome events are usually low in magnitude, often with emergent P-wave onsets, as well as P-waves shrouded in the coda of preceding events, during swarms. Such characteristics make the events difficult to identify using standard automatic detection and location procedures. We first evaluate current methods using an STA/LTA algorithm, coincidence event detectors, and pre-trained, deep-learning detectors and pickers. We find that detection of consistent P-wave phases across several stations for the same event is challenging and poses a major problem for event association and location. To address this problem, we initiate a manual review of all initial event associations to eliminate false positives that could incorrectly inflate the number of events in the catalog. We, therefore, developed a custom-trained detector and picker that outperformed other methods, and it identified multiple events that were recorded by >70% of the stations in the array. Our approach is well-suited for identifying events with emergent P-wave onsets and short durations (~2-10 s), and our method correctly identified a spike in seismicity in the days leading up to a well failure at the dome. Our methodology can be easily adapted for similar types of studies, such as volcano, mine and dam monitoring, and geothermal exploration.
more »
« less
This content will become publicly available on August 1, 2025
Deep Learning Forecasts Caldera Collapse Events at Kı̄lauea Volcano
Abstract During the 3 month long eruption of Kı̄lauea volcano, Hawaii in 2018, the pre‐existing summit caldera collapsed in over 60 quasi‐periodic failure events. The last 40 of these events, which generated Mw > 5 very long period (VLP) earthquakes, had inter‐event times between 0.8 and 2.2 days. These failure events offer a unique data set for testing methods for predicting earthquake recurrence based on locally recorded GPS, tilt, and seismicity data. In this work, we train a deep learning graph neural network (GNN) to predict the time‐to‐failure of the caldera collapse events using only a fraction of the data recorded at the start of each cycle. We find that the GNN generalizes to unseen data and can predict the time‐to‐failure to within a few hours using only 0.5 days of data, substantially improving upon a null model based only on inter‐event statistics. Predictions improve with increasing input data length, and are most accurate when using high‐SNR tilt‐meter data. Applying the trained GNN to synthetic data with different magma‐chamber pressure decay times predicts failure at a nearly constant stress threshold, revealing that the GNN is sensing the underling physics of caldera collapse. These findings demonstrate the predictability of caldera collapse sequences under well monitored conditions, and highlight the potential of machine learning methods for forecasting real world catastrophic events with limited training data.
more »
« less
- Award ID(s):
- 2040425
- PAR ID:
- 10574901
- Publisher / Repository:
- American Geophysical Union
- Date Published:
- Journal Name:
- Journal of Geophysical Research: Solid Earth
- Volume:
- 129
- Issue:
- 8
- ISSN:
- 2169-9313
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Abstract Event logs, comprising data on the occurrence of different types of events and associated times, are commonly collected during the operation of modern industrial machines and systems. It is widely believed that the rich information embedded in event logs can be used to predict the occurrence of critical events. In this paper, we propose a recurrent neural network model using time‐to‐event data from event logs not only to predict the time of the occurrence of a target event of interest, but also to interpret, from the trained model, significant events leading to the target event. To improve the performance of our model, sampling techniques and methods dealing with the censored data are utilized. The proposed model is tested on both simulated data and real‐world datasets. Through these comparison studies, we show that the deep learning approach can often achieve better prediction performance than the traditional statistical model, such as, the Cox proportional hazard model. The real‐world case study also shows that the model interpretation algorithm proposed in this work can reveal the underlying physical relationship among events.more » « less
-
The US Gulf Coast has several massive underground caverns within salt domes. These caverns can store vast amounts of hydrocarbons, including the US Strategic Petroleum Reserve, used to increase energy supplies during emergency shortages. Unstable caverns can collapse, leading to sinkhole formation and the release of gas. Previous studies have identified elevated seismicity and surface deformation as precursors to salt cavern collapse and sinkhole formation. However, identifying sporadic seismicity can be complicated, requiring complex methods for robust detection and characterization of events, especially in high-noise settings. We investigate deformation of the Sorrento salt dome in Louisiana using 8 months of data from seismic arrays first deployed in February 2020. Our arrays are comprised of 12 to 17 SmartSolo 3C seismic nodes, spaced 0.2-1.9 km and installed around the dome. We recorded more than 1.2 Tb of data, sampled at 500 Hz. Waveforms of identified events range from <1 s to over 30 s in length, rendering power detection methods like the STA/LTA inefficient. Building on recent studies that use machine learning methods to identify small magnitude (Mw -2.0 to 2.0) earthquakes, we developed a custom-trained convolutional neural network and applied it over sliding windows of the waveforms to detect earthquakes, pick P-wave arrivals, and reduce false positives. We correlated waveforms across all stations and identified events when they were observed on at least 60% of the array stations. We used spectrograms to infer fluid content around sources and to eliminate anthropogenic signals, including but not limited to, helicopters, trains, and boats from the catalog. Event locations were used to identify microearthquake swarms within the dome. Our preliminary results show elevated seismicity in the days preceding a well failure, suggesting our method can be used to monitor underground caverns and similar settings, such as mines, dams, and geothermal sites.more » « less
-
Abstract Inflationary deformation and very long period (VLP) earthquakes frequently accompany basaltic caldera collapses, yet current interpretations do not reflect physically consistent mechanisms. We present a lumped parameter model accounting for caldera block/magma momentum change, magma chamber pressurization, and ring fault (assumed vertical) shear stress drop. Pressurization of the underlying magma chamber is represented by a tri‐axial expansion source, and the combined caldera block/magma momentum change by a vertical single force. The model is applied to Kīlauea 2018 caldera collapse events, accurately predicting near field static/dynamic ground motions. In addition to the tri‐axial expansion source, the single force contributes significantly to the VLP waveforms. For an average collapse event with fully developed ring fault, Bayesian inversion constrains ring fault stress drop to ∼0.4 MPa and the pressure increase to ∼1.9 MPa. That the predictions fit both geodetic and seismic observations confirms that the model captures the dominant caldera collapse mechanisms.more » « less
-
Fault friction is central to understanding earthquakes, yet laboratory rock mechanics experiments are restricted to, at most, meter scale. Questions thus remain as to the applicability of measured frictional properties to faulting in situ. In particular, the slip-weakening distance d c strongly influences precursory slip during earthquake nucleation, but scales with fault roughness and is challenging to extrapolate to nature. The 2018 eruption of Kīlauea volcano, Hawaii, caused 62 repeatable collapse events in which the summit caldera dropped several meters, accompanied by M W 4.7 to 5.4 very long period (VLP) earthquakes. Collapses were exceptionally well recorded by global positioning system (GPS) and tilt instruments and represent unique natural kilometer-scale friction experiments. We model a piston collapsing into a magma reservoir. Pressure at the piston base and shear stress on its margin, governed by rate and state friction, balance its weight. Downward motion of the piston compresses the underlying magma, driving flow to the eruption. Monte Carlo estimation of unknowns validates laboratory friction parameters at the kilometer scale, including the magnitude of steady-state velocity weakening. The absence of accelerating precollapse deformation constrains d c to be ≤ 10 mm, potentially much less. These results support the use of laboratory friction laws and parameters for modeling earthquakes. We identify initial conditions and material and magma-system parameters that lead to episodic caldera collapse, revealing that small differences in eruptive vent elevation can lead to major differences in eruption volume and duration. Most historical basaltic caldera collapses were, at least partly, episodic, implying that the conditions for stick–slip derived here are commonly met in nature.more » « less
