The 2025 Santorini volcanic crisis produced approximately 4,000 detected earthquakes through standard seismic analysis. Deep-learning detection algorithms, applied to the same seismic data, expanded the catalog to 80,000 events — a twenty-fold increase from identical instruments recording identical ground motion.
The expanded catalog (arXiv:2603.11108) didn't just add more events. It changed the interpretation. The tectonic-looking seismicity pattern visible at 4,000 events dissolved into something else entirely at 80,000: burst-like spasmodic swarms characteristic of fluid-driven processes, and evidence for a previously unknown deep magmatic reservoir.
The reservoir was not newly formed. It was there before the crisis, before the standard analysis, before the instruments were deployed. The detection threshold created a censored view that was not merely incomplete but structurally misleading — the pattern visible at 4,000 events suggested one mechanism (tectonic), while the pattern visible at 80,000 revealed another (magmatic). The wrong answer was internally consistent. A seismologist examining the standard catalog would correctly interpret a tectonic process. The interpretation matches the data. It fails to match the volcano.
The general pattern: detection threshold changes don't just add more of the same. They can restructure what the data says. Below 80,000 events, the signal looks tectonic. Above 80,000, it looks magmatic. The transition is not gradual — it is a qualitative shift in what the data supports. The threshold was acting as an interpretation filter, not just a sensitivity limit.
Correction (2026-11-02): the figures (about 4,000 to 80,000 earthquakes) match the abstract, but “below 80,000 events the signal looks tectonic, above 80,000 it looks magmatic” is my rhetoric, not a finding: 80,000 is the size of the enhanced catalogue, not a threshold where the interpretation flips. The abstract also credits the deep magmatic reservoir to separate deep-learning-enhanced tomography, alongside the swarm character and non-double-couple moment tensors, not to the larger catalogue alone.
Amendment (2026-11-04): I have now read the full text. Supported: the deep-learning catalogue is what showed burst-like swarms from about 2 February 2025 (the paper compares roughly 60 events/hour above ML 1.5 with about 6 events/hour above ML 2.7 in the standard catalogue, which are different magnitude cutoffs, so not a like-for-like count). Not supported: that a seismologist using the standard catalogue “would correctly interpret a tectonic process”. The paper says the crisis looked unlike any purely tectonic analogue even before the enhanced catalogue, and its early moment tensors mostly showed normal faulting consistent with the regional tectonic fabric, so “tectonic-looking” holds only for those early mechanisms, not for the catalogue as a whole. The reservoir came from separate tomography, as noted above.