Before a volcano erupts, the ground trembles in a specific way. Seismic energy shifts from discrete earthquakes to continuous tremor. The frequency spectrum changes. Strain accumulates in patterns that, if you have enough data from that specific volcano, you can learn to recognize as precursors.
The problem is that most volcanoes do not have enough data. Of roughly 1,500 potentially active volcanoes on Earth, only a handful are instrumented with decades of seismic monitoring. The rest have sparse records, or none at all. A machine learning model trained on Kīlauea cannot be naively applied to Merapi, because every volcano has its own geology, plumbing system, and eruptive style.
Or so the assumption went. Dempsey and colleagues tested the assumption by asking whether seismic precursors are ergodic — whether the distribution of pre-eruptive signals across an ensemble of different volcanoes looks like the distribution over time at any single volcano. They assembled 41 eruptions at 24 volcanoes spanning 73 years and found that it does. The ensemble statistics serve as a proxy for the temporal statistics at any individual site.
This is not an obvious result. Ergodicity requires that the underlying process is, in some meaningful sense, the same everywhere. For earthquakes along a fault, the Gutenberg-Richter distribution provides this universality. But volcanoes differ enormously in magma composition, conduit geometry, and eruption mechanism. The ergodic property means that beneath this variation, the mechanics of pressure accumulation and failure follow shared statistical laws. The individual volcano is a single realization of a common process.
The practical consequence is striking: a transfer learning model trained on the ensemble — with all data from the target volcano withheld — forecasts eruptions at the unseen volcano with accuracy comparable to a model trained directly on that volcano's own data. The ensemble knows what the individual volcano has not yet shown.
The through-claim is about what universality costs and what it buys. Treating each volcano as unique respects geological specificity but leaves most volcanoes unforecastable. Treating all volcanoes as instances of a common process sacrifices specificity but extends forecasting to the 1,500 sites that have no data of their own. The ergodic finding means the universality is not imposed — it is measured. The volcanoes share a statistical language of failure whether we model them individually or not. The ensemble does not erase the individual. It reveals that the individual was always a member.