Deep Learning Forecasts Caldera Collapse Events at K\=ilauea Volcano

• DOI: 10.48550/arxiv.2404.19351
• Publication Date: 2024-04-30
• Authors: Ian W. McBrearty; P. Segall
• Keywords: Caldera

During the three month long eruption of K\=ilauea volcano, Hawaii in 2018, pre-existing summit caldera collapsed over 60 quasi-periodic failure events. The last 40 these events, which generated Mw >5 very period (VLP) earthquakes, had inter-event times between 0.8 - 2.2 days. These events offer a unique dataset for testing methods predicting earthquake recurrence based on locally recorded GPS, tilt, and seismicity data. In this work, we train deep learning graph neural network (GNN) to predict time-to-failure collapse using only fraction data at start each cycle. We find that GNN generalizes unseen can within few hours 0.5 days data, substantially improving upon null model statistics. Predictions improve with increasing input length, are most accurate when high-SNR tilt-meter Applying trained synthetic different magma pressure decay predicts nearly constant stress threshold, revealing is sensing underling physics collapse. findings demonstrate predictability sequences under well monitored conditions, highlight potential machine forecasting real world catastrophic limited training