Toward Operational Probabilistic Eruption Forecasting Using Machine Learning and Seismic Tremor: A Retrospective Study of the 2023 Shishaldin Paroxysms (Alaska)

Rapid and accurate forecasting of volcanic eruptions remains a central challenge for volcano surveillance agencies. Traditionally, forecasting efforts have focused on recognizing recurrent patterns in geophysical or geochemical signals to detect unrest and assess its evolution; however, translating these precursory signals into clear, easy-to-interpret eruption probabilities remains challenging. A promising signal in the context of probabilistic eruption forecasting is seismic tremor, as it often exhibits recognizable patterns (e.g., amplitude escalation, frequency shifts, and spectral variations) during the run-up to eruptions. This raises the following question: Can seismic tremor patterns be used operationally to produce objective eruption probabilities? To address this question, we developed a supervised machine learning-based framework built upon the Dempsey et al., 2020 [https://doi.org/10.1038/s41467-020-17375-2], Ardid et al., 2023 [https://doi.org/10.21203/rs.3.rs-3483573/v1], and Girona and Drymoni, 2024 [https://doi.org/10.1038/s41467-024-51596-z] approaches, and tested it retrospectively on 13 paroxysmal events at Shishaldin Volcano, Alaska, that occurred between July and November 2023.  Specifically, our framework extracts statistical features from continuous tremor data, such as dominant frequency, amplitude, kurtosis and Shannon entropy, and applies a Random Forest classifier to quantify the similarity between ongoing tremor and previously recorded pre-eruptive tremor; this similarity can, in turn, be interpreted as an estimate of the probability of an eruption occurring within a specific time window. To mimic operational conditions, models were retrained on progressively larger datasets, using only data available prior to each Shishaldin paroxysm; and forecasts targeted seismic amplitude peaks and the onset of ash emissions for 1, 6, 12, and 24-hour windows. Results show that, in most cases, probabilities increased in the lead-up to the paroxysms, indicating that our approach captured evolving tremor patterns associated with imminent explosive activity. Although evaluated retrospectively, the findings highlight the potential of seismic tremor–based probabilistic forecasts to support volcano monitoring and decision-making during volcanic crises. The framework is fully retrainable, automatically updating as new paroxysms occur and additional data become available, thereby enhancing its suitability for near-real-time operational use and enabling straightforward extension to other volcanic systems.