Cross-Volcano Eruption Timing from Seismicity Alone

Operational volcano alert systems issue categorical warnings — Normal, Advisory, Watch, Warning — based on expert judgment. They work, but they don't answer the question that matters most during a crisis: how many days until eruption? No existing operational system produces a quantitative timing estimate that generalizes across volcano types without per-volcano recalibration. We built one.

The question

Is there a universal timing signal in pre-eruptive seismicity that can be learned from one set of volcanoes and applied to others? Cross-volcano generalization is the operational test that matters. At an unmonitored or newly active volcano, per-volcano training data does not exist. If the timing signal is volcano-specific, the system is useless where it's needed most. If it transfers, it could change how eruption warnings work.

Three volcanoes, three eruption styles

We trained on publicly available FDSN seismicity catalogs from three volcanoes chosen for their diversity. Kilauea is a basaltic shield volcano in Hawaii with frequent effusive eruptions — five training eruptions from 2008 to 2021, with 140 to 2,159 seismic events per 60-day pre-eruption window. Piton de la Fournaise is an ocean-island basaltic volcano on Reunion Island — five eruptions from 2020 to 2023, with 89 to 159 events per window. Redoubt is an arc explosive volcano in Alaska — one eruption in 2009, with only 21 events in the sparse public catalog. These are structurally different volcanoes with different magma compositions, eruption styles, and seismic signatures.

The system extracts ten features from each 60-day pre-eruption seismicity window: persistence and count of volcano-tectonic events, acceleration of seismicity rate, depth migration slope, fraction of long-period events, azimuthal asymmetry, radial deformation rate and acceleration from continuous GPS, and inter-eruption repose fraction. A ridge regression model estimates days-to-eruption. A logistic regression model estimates the probability that an eruption will occur within 7 days. A hybrid decision rule combines both with convergence scoring and persistence gates.

The results

We evaluated under two protocols. Leave-one-eruption-out (LOEO) holds out one eruption at a time from the same volcanoes — the standard within-dataset test. Leave-one-volcano-out (LOVO) holds out an entire volcano — the operational generalization test, where the system has never seen any eruption from that volcano type.

LOVO timing error at 7 days before eruption: 9.6 ± 3.0 days. LOEO timing error: 8.7 ± 5.7 days. The difference is 0.9 days. The timing signal learned from basaltic shield and ocean-island volcanoes transfers to arc explosive volcanoes — and vice versa — with almost no degradation. Mean alert lead time is 29 days in LOVO and 24 days in LOEO.

False alert rate: 0.0% across 1,551 quiet test days spanning five baselines. This includes Kilauea during its continuously active Pu'u O'o period (1,277 seismic events, zero false alerts), two Redoubt post-eruptive years, and — critically — Mauna Loa in 2010, a volcano the system had never seen during training, 26 years into an inter-eruptive period. Zero false alerts on an unseen volcano is the strongest evidence that the system has learned a real timing signal rather than overfitting to training volcanoes.

Training Volcanoes and Eruptions

False Alert Validation

The Piton 2023 quiet period is particularly informative: a seismicity burst in January produced high alert probabilities (P = 0.58–0.66) but low multi-channel convergence (0.02–0.14). The convergence gate correctly suppressed 6 potential false alerts.

The holdout test

The 2018 Kilauea Lower East Rift Zone eruption was reserved as a fully held-out case — never used in any training or cross-validation fold. This was the most significant Hawaiian eruption in decades, destroying over 700 structures and reshaping the Puna coastline. The system's timing estimate at 7 days before onset had a mean absolute error of 9.5 days. At 14 days before onset, the error dropped to 0.6 days. The first alert fired at 7 days pre-eruption via the hybrid rule channel.

What transfers and what doesn't

The 0.9-day gap between LOVO and LOEO timing error is the key result. It means the pre-eruptive seismicity signal is not volcano-specific — a timing model trained on Hawaiian and Reunion eruptions produces useful estimates for an Alaskan arc volcano it has never seen, and vice versa. This supports the hypothesis that pre-eruptive seismicity carries a universal timing signal learnable from catalog-quality data.

The signal appears to be in the acceleration and persistence of volcano-tectonic events, the migration of earthquake depths, and the evolving ratio of long-period to volcano-tectonic events — features that reflect the physics of magma ascent regardless of the specific volcanic setting. The details differ (Kilauea produces thousands of events, Redoubt produces dozens), but the temporal pattern of approach is consistent.

Honest limitations

Three volcanoes is a small training set. We need at least four for robust leave-one-volcano-out statistics. Redoubt's sparse catalog (21 events) inflates the k=14 timing error. Two volcanoes we evaluated for inclusion — Pavlof and Etna — were excluded because their public seismicity catalogs are too incomplete at low magnitudes. Pavlof's pre-eruptive swarms are below M1.0 and absent from public FDSN endpoints entirely. Etna's public catalog (M ≥ 1.5 only) produces pre-eruptive records indistinguishable from background, degrading training quality from 11.1 to 92.2 days MAE.

SO2 satellite data (TROPOMI) has not yet been integrated as a third convergence channel. Adding it should improve both the convergence gate (fewer near-miss false alerts) and the timing estimate (gas flux carries independent information about magma proximity to the surface). That work is underway.

All training and evaluation was performed on local hardware using publicly available data. No proprietary monitoring networks, no cloud compute, no per-query API costs.

Working on volcanic monitoring, eruption forecasting, or seismicity analysis? Reach us at trevin@lytelab.ai