Forecasting volcanic eruptions
Unrest at a volcano is often visible, but when an eruption starts, how big it gets and what kind it is are still mostly unpredictable.
Open in the interactive tree →Observatories can tell when a volcano is restless, and at a few closely watched ones they can give a window of days for the next event. What stays unreliable is whether unrest ends in an eruption at all, and how large and explosive it will be. USGS counts about 1,350 potentially active volcanoes, and a 2018 Bristol-led preprint estimated that around half of roughly 1,500 active volcanoes have no ground-based monitoring.
As of October 2026
October 2026: at Kilauea (Hawaii) the episodic lava-fountaining eruption that began in December 2024 reached episode 54 on 25 August 2026; HVO raised the alert back to WATCH on 7 September, new vents opened on the crater floor on 14-16 September, and a swarm of 80+ earthquakes with rapid inflation on 30 September-1 October points to another episode, but no fountaining had occurred since 25 August. In Iceland the Met Office estimates about 29.1 million cubic metres of magma beneath Svartsengi (16 August 2026), but warning times before earlier events ran only from about 20 minutes to just over four hours. At Campi Flegrei near Naples the ground rises about 10 mm a month (31 cm since January 2025 at one station, excluding the jump from a strong earthquake on 31 July), and INGV reports no significant change in the deformation pattern; at Hunga Tonga, declared dormant on 11 January 2022, the climactic eruption followed on 15 January with a column about 57-58 km high. A February 2025 Nature Communications study showed that seismic precursors learned at some volcanoes carry over to unseen ones.
What is missing
- Signs that separate unrest that ends in eruption from unrest that fades, as at Campi Flegrei, where uplift continues
- Forecasts of eruption size and style, not only of the start
- Ground instruments on the roughly half of active volcanoes that have none, and monitoring of underwater vents like Hunga Tonga's
- Enough well-recorded eruptions per volcano to train and test models, since many volcanoes erupt only rarely
- Real-time operational tests of cross-volcano models, not only back-tests on past eruptions
Becomes possible once solved
- Evacuations and flight rerouting timed days ahead instead of hours
- Fewer needless evacuations around restless volcanoes such as Campi Flegrei
- Land-use, insurance and aviation planning based on expected eruption size
Open steps
- Cross-volcano precursor models High AI leverageTest whether seismic precursors learned at some volcanoes transfer to unseen ones in real time, including steam-driven eruptions with little warning.
- Global radar deformation screening High AI leverageRun automated detection of ground inflation and deflation on satellite radar data for every volcano, and flag unrest at ones with no ground instruments.
- Will unrest end in eruption? Medium AI leverageTell unrest that ends in eruption from unrest that fades, as at Campi Flegrei (31 cm uplift at one station since January 2025) and on Iceland's Reykjanes peninsula.
- Predicting eruption size and style Medium AI leverageEstimate from unrest data how large an eruption will be and whether it will be effusive or explosive, as in Hawaii's lava fountains versus Hunga Tonga's 57 km column.
- Monitoring underwater volcanoes Low AI leverageDetect precursors at submarine vents from hydroacoustic, remote seismic and satellite data; Hunga Tonga was declared dormant four days before its 2022 blast.
Where AI could help
Medium AI leverage. AI finds precursors in seismic and satellite data across volcanoes, but eruptions are rare and many volcanoes lack instruments; size and style stay unpredicted.
- Find recurring seismic, gas and deformation precursors across many volcanoes
- Scan satellite radar images of all volcanoes for ground movement
- Combine monitoring streams into probabilistic forecasts with honest error bars
- Detect and classify small quakes and tremor in continuous data
Shown so far
- In February 2025, a Nature Communications study trained a transfer-learning model on seismic data from 41 eruptions at 24 volcanoes over 73 years; on volcanoes left out of training it matched locally trained models and beat seismic-amplitude benchmarks. source
- In January 2018, a Bristol-led preprint used convolutional neural networks to flag ground deformation in satellite radar images, because manual inspection of every volcano is impractical. source