Deep Learning Forecasts Caldera Collapse Events at Kı̄lauea Volcano.

During the 3 month long eruption of Kı̄lauea volcano, Hawaii in 2018, the pre‐existing summit caldera collapsed in over 60 quasi‐periodic failure events. The last 40 of these events, which generated Mw > 5 very long period (VLP) earthquakes, had inter‐event times between 0.8 and 2.2 days. These failure events offer a unique data set for testing methods for predicting earthquake recurrence based on locally recorded GPS, tilt, and seismicity data. In this work, we train a deep learning graph neural network (GNN) to predict the time‐to‐failure of the caldera collapse events using only a fraction of the data recorded at the start of each cycle. We find that the GNN generalizes to unseen data and can predict the time‐to‐failure to within a few hours using only 0.5 days of data, substantially improving upon a null model based only on inter‐event statistics. Predictions improve with increasing input data length, and are most accurate when using high‐SNR tilt‐meter data. Applying the trained GNN to synthetic data with different magma‐chamber pressure decay times predicts failure at a nearly constant stress threshold, revealing that the GNN is sensing the underling physics of caldera collapse. These findings demonstrate the predictability of caldera collapse sequences under well monitored conditions, and highlight the potential of machine learning methods for forecasting real world catastrophic events with limited training data. Plain Language Summary: During the summer of 2018, Kı̄lauea volcano, Hawaii, experienced a dramatic series of large earthquakes, coinciding with the collapse of the summit caldera in a series of repeated failure events. These collapse events occurred periodically, with inter‐event timings between 0.8 and 2.2 days. Because of the significance of this event, there is interest to understand more about the dynamics of this collapse sequence. We study whether observational measurements of deformation recorded on the surface of the volcano carry signatures that indicate the timing of the upcoming collapse events. By using machine learning, we train a series of models that aim to predict the time‐to‐failure of each cycle based on the observed deformation data, and we experiment with using different combinations of input data sets. We find our models can accurately predict the timing of most collapse events to within a few hours, including for events that the models were never trained on and that have longer durations than the training events. These results shed new light on the dynamics and predictability of the Kı̄lauea caldera collapse sequence. Key Points: A sequence of >M5 caldera collapse earthquakes occurred at Kı̄lauea Volcano with recurrence intervals of 0.8–2.2 daysA Graph Neural Network trained with locally recorded GPS and tilt data could predict failure times of most collapse events to within a few hoursThe deep learning model generalizes to cycles longer than the training cycles and can handle arbitrarily long time series inputs [ABSTRACT FROM AUTHOR]