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

The 2014 phreatic eruption at Mt Ontake, Japan, occurred without clear surface precursors and caused significant casualties, underscoring the difficulty of forecasting non-magmatic (steam-driven) eruptions. Prior retrospective studies found useful signals only seconds to minutes before eruption or long-term changes from sparse data, while mid-term geodetic and seismic precursors remained ambiguous. Anticipating such eruptions requires integrating new methodologies. Recent advances in volcano monitoring include seismic interferometry for detecting tiny seismic velocity changes and machine learning for automatic classification and detection of seismo-volcanic signals. This study aims to determine whether previously overlooked precursory signals existed months before the 2014 eruption by applying single-station seismic interferometry, network-based coherence analyses, and a pre-trained deep-learning earthquake detector to continuous data from the Mt Ontake network.

Previous work on phreatic eruptions identified short-term (seconds–minutes) precursors (e.g., Kato et al.) too late for actionable warning, and long-term indicators (e.g., water chemistry, ground level; Sano et al.) that are temporally sparse and less useful for real-time forecasting. Mid-term geodetic uplift prior to Ontake 2014 is ambiguous, and stress field deviations inferred from seismicity are challenging to operationalize and require sustained earthquake activity. Seismic interferometry has been used to detect subtle crustal velocity changes before and during eruptions (e.g., Brenguier et al.; Donaldson et al.), while machine learning has enabled supervised and unsupervised classification of seismic signals and robust phase picking (e.g., PhaseNet, EQTransformer). These advances motivate re-examining continuous Ontake records for early, subtle signals related to hydrothermal processes preceding phreatic activity.

Data and preprocessing: Continuous seismic data from 11 stations (Nagoya Univ., JMA, Gifu Prefecture, NIED) were corrected for timing and gaps, band-pass filtered 0.01–8 Hz, and resampled to 20 Hz. Autocorrelations (ACF) were computed primarily on the vertical component (horizontal and cross-component correlations were too scattered). Signals were clipped at 3 rms and stacked over 5-day windows to balance precision and temporal resolution. Seismic velocity changes dv/v were estimated by comparing daily AC functions to a full-period reference using Moving Window Cross Spectral Analysis (MWCS) over ±5 to +35 s coda windows, assuming homogeneous velocity changes. A Brenguier-style linear inversion for continuous dv/v time series was also tested (weighting 100,000) with similar results, albeit with higher computation time. Robustness checks included waveform stretching and Phase Cross-Correlation (PCC), which reproduced the key dv/v pattern. Seasonal and environmental correction: To mitigate non-volcanic perturbations, dv/v time series were corrected for pore pressure effects from precipitation and snow depth (following Wang et al.), using daily precipitation, snow data, and a diffusion model (c=1 m^2/s) within r=8 km. Atmospheric pressure effects were assessed; dv/v-to-load coefficients were on the order of 0.1–1%/m, consistent with prior studies. The dv/v series was smoothed with a Hodrick–Prescott filter (factor 10,000) for trend visualization. Sensitivity kernels and spatial constraints: Phase-velocity sensitivity kernels for Rayleigh and Love waves were computed (surf96) using regional 1D models, indicating high sensitivity at depths 0–1.5(–2) km beneath summit stations in the 1–2 Hz band. Additional 2D kernels accounting for diffusion and absorption (velocity 1 km/s, mean free path 1 km, 15 s coda, 500 m grid) further constrained sensitivity local to summit vicinity. Cross-station CCFs did not consistently show the effect due to attenuation and loss of coherence with distance, while single-station ACF and PCC did. Network-based tremor detection: A covariance-matrix eigenvalue analysis (spectral width proxy for source coherence) was applied to 0.5–15 Hz band-passed data. Traces were decimated to 50 Hz, 1-bit normalized per station, segmented into 50% overlapping windows and 100 s subwindows. Low spectral width indicated coherent tremor episodes. Deep-learning earthquake detection: The pre-trained EQTransformer was run in transfer mode on continuous data (stations KID and ONTA) using low detection/picking thresholds. Detections were visually vetted for plausibility. S–P times were used qualitatively to infer source distances. Complementary observations: Groundwater pressure from the GOT well (sealed gauge at ~−650 m depth, ~10 km SE of summit) was compared with dv/v. Reported station issues in July–August 2014 were noted. Published geodetic and seismic observations from prior studies were reviewed for correlation with dv/v. Modeling: A rainfall-driven pore-pressure model and a linear snow-load relationship were used to synthesize environmental dv/v contributions. Volumetric strain sensitivity to groundwater level changes was used to estimate strain equivalents. Numerical context for inflation-induced volumetric strain distributions was considered to interpret spatial dv/v patterns.

- A localized dv/v decrease was detected at summit stations in the 1–2 Hz band from June to October 2014, with continued decrease for a few days post-eruption. Lower frequency bands (≤1 Hz) did not show this change, consistent with deeper sampling and higher confining pressure. - The dv/v perturbations were confined near summit stations and sensitive at depths ~0–1.5 km beneath the surface, indicating shallow structural changes likely within the volcano-hydrothermal system. - After environmental correction, dv/v exhibited a long-term increase from May 2013 to late April 2014 (interpreted as sealing/reduced permeability), followed by fluctuations from May to mid-August 2014 during a period previously considered quiescent, and a significant decrease (~−0.04%) coincident with the onset of volcano-seismicity in late August. - dv/v variations correlate with volumetric strain inferred from groundwater pressure at the GOT well (~10 km SE): a ~20 cm groundwater level drop corresponds to ~95 nstrain increase in volumetric strain, matching dv/v decreases, suggesting cycles of pressurization/depressurization in April–August 2014. - The network-based analysis revealed a coherent tremor episode between 1.3–4.0 Hz from 15–18 August 2014, the only such occurrence in May–August, preceding the final dv/v reduction by ~2 days. Another tremor episode accompanied the 27 September eruption. - EQTransformer detected 3265 earthquakes. Regional bursts occurred on 3 May 2014 (near Yakeda, ~45 km away, M up to ~4) coincident with the onset of dv/v decrease, and on 22 July 2014 (~30 km north; S–P 3.2–3.6 s; M up to ~4) followed by dv/v decreases. dv/v sensitivity to distal earthquakes suggests critically pressurized fluids beneath Ontake’s SE flank. - No large regional earthquakes (M≥4 within 100 km) immediately prior to the eruption were reported; dv/v changes are thus not attributable to major regional events. - Post-eruption, correlation coefficients between daily and reference ACFs recovered to pre-eruptive levels by December 2014, implying no extensive permanent damage in the shallow edifice. - Conceptual staging: Stage 1 (May 2013–Apr 2014) sealing/increasing dv/v; Stage 2 (Apr–Aug 2014) cycles of pressurization/strain with dv/v fluctuations and sensitivity to distal quakes; Stage 3 (Aug–Sep 2014) over-pressurization leading to inflation, local seismicity, and eruption initiation.

The identified sequence of shallow dv/v changes months before the 2014 phreatic eruption indicates the presence and evolution of pressurized hydrothermal fluids beneath Mt Ontake’s summit, particularly the SE flank. The long-term dv/v increase likely reflects sealing and reduced permeability, heightening susceptibility to pressure build-up. Subsequent dv/v fluctuations and correlation with volumetric strain and groundwater pressure suggest cycles of pressurization and depressurization starting as early as April–May 2014, despite apparent quiescence in conventional observables. The dv/v response to distal earthquakes in May and July indicates a critically stressed, fluid-rich system susceptible to dynamic stress perturbations. The mid-August coherent tremor, followed by dv/v decrease and later uptick in local seismicity and slight inflation in late August, is consistent with over-pressurization culminating in the 27 September eruption. The localization of dv/v changes to shallow summit regions aligns with surface-wave sensitivity and numerical expectations of near-surface strain amplification during inflation. Collectively, these observations show dv/v as a sensitive proxy for tracking pressurized volumes and poroelastic effects in hydrothermal systems, addressing the challenge of forecasting phreatic eruptions by revealing subtle, early-stage processes months before failure.

This study reveals previously hidden precursory processes before the 2014 Mt Ontake phreatic eruption: a year-long dv/v increase (sealing), followed by multi-month cycles of pressurization evidenced by dv/v–strain correlations and sensitivity to distal earthquakes, and finally over-pressurization leading to local seismicity, slight inflation, tremor, and eruption. Single-station seismic interferometry at summit stations captured shallow (0–2 km) dv/v changes beginning about five months pre-eruption, demonstrating its potential for early detection of non-magmatic unrest. Incorporating dv/v with network-based coherent signal detection, groundwater/strain observations, and machine learning-based detection can enhance real-time monitoring and forecasting. Future work should deploy instruments near summits and hydrothermal areas to improve sensitivity, integrate dv/v into ML forecasting frameworks (e.g., transfer learning from Whakaari models), perform comprehensive cross-volcano studies of pre-phreatic dv/v behavior, and develop improved modeling to discriminate top-down sealing from bottom-up magmatic gas influx.

- Spatial sensitivity is limited at higher frequencies (>1 Hz), restricting dv/v detection largely to summit-proximal shallow volumes; cross-station CCFs often lost coherence with distance, limiting traditional interferometry. - The seismic network geometry prevented location of the mid-August tremor source; attribution remains uncertain. - Seasonal and atmospheric pressure effects require careful correction; residual environmental influences and short time series complicate absolute dv/v interpretation. Some abnormal January–March 2014 fluctuations coincided with high atmospheric pressure. - Post-eruption (Oct–Dec 2014) coherence was low, hindering reliable dv/v estimation; only by December did coherence recover. - The GOT well experienced sensor issues (late July–October 2014), limiting high-frequency comparison with dv/v during that time. - Raw seismic data are not publicly available due to privacy laws, potentially limiting reproducibility; only derived ACFs are archived. - Averaging across stations may obscure localized effects; care is needed in network-level summaries.