Array-wise self-supervised encoding of the volcanic wavefield tracks the continuum of an eruption: the 2025 to 2026 reawakening of Piton de la Fournaise volcano
Quentin Higueret, Zachary E. Ross, & Yaozhong ShiSubmitted August 30, 2026, SCEC Contribution #15226, 2026 SCEC Annual Meeting Poster #TBD
Volcano observatories track unrest by compressing the continuous seismic wavefield into a few hand-crafted summaries such as the real-time amplitude, the earthquake catalog and the coherence of the network. Each reduces the record to one property, tuned to one kind of signal and read one station at a time, so most of what is recorded goes unused. In this study, we investigate how much of an eruption can be recovered purely from raw array data without ground-truth labels or engineered features. To achieve this, we train a neural network to predict the spectrogram of an unseen station using only the surrounding seismic stations. Because it must fill in a hidden station, the network is forced to learn how the energy is organized across the whole array rather than within a single trace. The encoder is told only where each station sits, never which station it is, so one set of weights applies to any geometry. We train it on the 2018 to 2019 activity of Piton de la Fournaise, La Reunion, then freeze it and apply it without retraining to the 2025 to 2026 reawakening: a transfer test in both time and array geometry on the same volcano. Clustering the frozen representation recovers the reactivation without labels, from the deep recharge through the successive flank eruptions, the pauses and the closing gas-piston tremor, and its leading coordinate closely tracks an independent network-coherence chronology computed daily. Because the input is amplitude-normalized, this cannot be a restatement of amplitude. The eruption reads as a smooth continuum of states rather than separable event types, yet at finer scale the representation resolves sub-states the operational catalog merges into its background. One frozen representation serves several monitoring analyses at once, pointing to learned, array-wise wavefield encoders as a reusable, label-free basis for volcano monitoring.
Key Words
volcano seismology,wavefield,eruption, neural operator
Citation
Higueret, Q., Ross, Z. E., & Shi, Y. (2026, 08). Array-wise self-supervised encoding of the volcanic wavefield tracks the continuum of an eruption: the 2025 to 2026 reawakening of Piton de la Fournaise volcano. Poster Presentation at 2026 SCEC Annual Meeting.
Related Projects & Working Groups
Seismology
