Volcanic eruption forecasting has a data scarcity problem. Most volcanoes erupt rarely — decades to centuries between events. A well-monitored volcano might have seismic data covering one or two eruptions. Training a machine learning model on one or two positive examples is a recipe for overfitting.
Transfer learning across volcanoes (Nature Communications, 2025) solves this by identifying “ergodic” seismic precursors — signal features that change consistently before eruptions across different volcanoes. Using data from 41 eruptions at 24 volcanoes over 73 years, the study finds common precursor patterns transferable between sites: certain frequency ratios, amplitude trends, and event rate accelerations that precede eruptions regardless of the volcano's specific magmatic system.
The counterintuitive implication: a model trained on volcanoes you've never visited may predict eruptions at your volcano better than a model trained on your volcano's sparse local data. The transferred knowledge from 23 other eruption sequences provides a richer representation of pre-eruptive behavior than the one or two sequences your own seismic network recorded.
This works because the physics of magma movement shares common features across eruption types. Rising magma fractures rock (producing seismicity), pressurizes gas (changing frequency content), and deforms the edifice (changing wave propagation). The details differ — basaltic vs. silicic, phreatic vs. magmatic — but the qualitative patterns of increasing rate, changing spectral content, and accelerating deformation are common enough to transfer.
The approach echoes medical diagnostics: a model trained on thousands of patients performs better than a model trained on one patient's history, even though patients differ. The population-level pattern captures what the individual case cannot. For volcanoes as for patients, individual history is useful but insufficient; cross-system experience is where statistical power lives.
Note (2026-11-16; checked against the open-access Nature Communications paper, 'Ergodic seismic precursors and transfer learning for short term eruption forecasting at data scarce volcanoes', as summarized from the full text by a page fetch, plus its abstract): the 41 eruptions at 24 volcanoes and the finding that precursors share patterns across volcanoes are the paper's. On performance, the paper says tailored (target-volcano) and generalized models have 'comparable discriminability, as indicated by similar AUC values' (about 0.75 to 0.96 for the volcanoes it highlights), which is comparable, not better, so the essay's 'a model trained on volcanoes you've never visited may predict eruptions at your volcano better than a model trained on your volcano's sparse local data' goes beyond it. The summary also reports that the paper does not claim the patterns hold regardless of magmatic system: it trains separate models for magmatic and phreatic eruptions, the global model performing worse than specialized ones, and it lists limits (a 48-hour forecast window, high false-positive rates at some volcanoes such as Mt. Ontake, poor performance on a closed-conduit fissure eruption at Eyjafjallajökull, and no sensitivity to some precursors before Cordón Caulle 2011). The precursor measures it names include high- and median-frequency bands, RSAM amplitude and a median-to-high-frequency ratio (DSAR); I have not checked the essay's list of 'frequency ratios, amplitude trends and event rate accelerations' against the text. The essay's account of why it works (magma fracturing rock, pressurizing gas, deforming the edifice) is its own.