Hidden pressurized fluids prior to the 2014 phreatic eruption at Mt Ontake

The 2014 Mt. Ontake eruption, a phreatic eruption resulting in over 50 deaths, underscored the limitations in predicting such non-magmatic events. These eruptions, while frequent, are challenging to forecast due to the lack of clear precursor signals. Previous studies examining this eruption have identified short-term precursors (seconds to minutes before) and long-term changes (years), but these are either too late for effective warnings or too sparse for real-time forecasting. Some mid-term precursors (days to one month) have been suggested from geodetic and seismic data, but these remain ambiguous and difficult to implement widely. This research addresses this gap by employing advanced seismic interferometry and deep learning techniques to analyze continuous seismic data, searching for previously undetected precursors to the Mt. Ontake eruption.

Several studies have explored the use of seismic interferometry to monitor volcanoes, successfully detecting changes in seismic wave propagation before and during magmatic eruptions. Machine learning has also emerged as a powerful tool for classifying seismo-volcanic signals, with supervised and unsupervised approaches being developed. However, the application of these advanced techniques to the Mt. Ontake eruption specifically, to identify potential precursory signals during the seemingly quiescent period before the eruption, remains largely unexplored. This study builds upon this existing work by integrating the latest advancements in both seismic interferometry and deep learning for a more comprehensive analysis.

The study utilized daily single-station inter-component auto-correlations (ACF) and inter-station single-component cross-correlations (CCF) from 11 seismic sensors at Mt. Ontake to estimate seismic velocity changes. Seismic velocity variations (dv/v) were calculated from 5-day stacks of ACF, compared to a full-period stack as a reference. The researchers employed various processing parameters and data processing approaches to ensure robustness. To account for non-volcanic perturbations, they mitigated seasonal effects by computing pore pressure changes from daily precipitation and snow depth. The approach of Brenguier et al. (2014) was used for more stable velocity change estimates. A network-based method was implemented for detecting volcanic tremor by analyzing eigenvalues of the seismic network covariance matrix. The EQTransformer deep-learning model was used in a predictive mode on continuous data to detect earthquakes. Finally, complementary observations, such as groundwater pressure changes, were analyzed for correlations with seismic velocity variations.

The study revealed a significant pre-eruptive drop in seismic velocity (dv/v) from June to October 2014 at sensors located on the summit, specifically in the 1–2 Hz frequency band. This velocity drop was spatially localized, observed only at summit stations and within this specific frequency range. The most significant velocity reduction coincided with slight crustal deformation in mid-August, followed by volcano-earthquakes and stress changes. A network-based analysis detected a coherent tremor between 1.3–4.0 Hz from August 15-18, the only such event during the May-August 2014 period. The EQTransformer model detected two periods of enhanced seismicity, correlating with velocity drops. A striking correlation was found between the dv/v time-series and groundwater pressure changes at a well 10 km from the summit. Analysis suggests three stages: Stage 1 (May 2013 to April 2014): sealing of the subsurface; Stage 2 (April 2014 to August 2014): pressurization cycles of extension-compression; and Stage 3 (August 2014 to September 2014): over-pressurization, culminating in the eruption. The seismic velocity decrease continued for a few days after the eruption but recovered to pre-eruptive levels by December 2014.

The findings suggest that the 2014 Mt. Ontake eruption was likely triggered by the progressive ingress of gas/steam accumulating below the eastern crater at shallow depths, creating critically stressed conditions. This accumulation started at least five months before the eruption. The correlation between seismic velocity, volumetric strain, and groundwater pressure strongly indicates the involvement of pressurized fluids. The localized nature of the seismic velocity changes suggests that monitoring near the summit and within the volcano-hydrothermal system is crucial. The reversible nature of the stress changes, inferred from the seismic velocity variations, suggests a cyclical process of pressurization that could potentially recur.

This study demonstrates that single-station seismic interferometry, combined with deep learning and other observational data, can effectively detect slow, early-onset precursory signals for phreatic eruptions. The detected pre-eruptive changes in seismic velocity, strongly correlated with groundwater pressure changes, provide valuable insights into the mechanisms leading up to the 2014 eruption. Future research should focus on expanding the use of these techniques to other volcanoes, potentially incorporating dv/v and other parameters into machine learning models for enhanced real-time monitoring and prediction of phreatic eruptions.

The spatial sensitivity of the seismic velocity changes at higher frequencies (above 1 Hz) is limited. The study relies on retrospective analysis and data availability limitations from the event. While seasonal variations were mitigated, some uncertainties might still remain. The lack of extensive microseismicity preceding the major velocity drop in August might indicate a limitation in resolving smaller-scale precursory signals.