Deep Learning Forecasts Caldera Collapse Events at Kilauea Volcano

TL;DR

The paper demonstrates that a Graph Neural Network operating on temporally windowed GPS, tilt, and seismicity data can forecast Kīlauea caldera-collapse times within a few hours using only early-cycle observations ($T$), achieving $R^2$ exceeding $0.95$ in the final 24 hours. The model integrates local and global information via a GraphSAGE-like convolution on temporal graphs, with time-embedding $\boldsymbol{t}_i$ and edge features $\boldsymbol{e}_{ij}=\boldsymbol{h}_i-\boldsymbol{h}_j$, and is trained on the first $29$ cycles and validated on the last $10$. Results show multi-modal inputs (GPS+tilt) substantially outperform single-sensor or non-ML baselines, and the approach generalizes to longer-than-training cycles, indicating extrapolative capability. The work highlights the potential for ML-enabled real-time forecasting of catastrophic geophysical events under well-instrumented conditions, and outlines future extensions to spatial-temporal graphs, domain adaptation, and application to other volcanoes and fault systems.

Abstract

During the three month long eruption of Kilauea 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 - 2.2 days. These failure events offer a unique dataset 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 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.

Figures (6)

• Figure 1: Cumulative seismicity (a) and radial displacement at GPS station CRIM (b) during the final 17 caldera collapse cycles of the 2018 Kilauea eruption. The red portion of each trace in (a,b) indicates an example 0.5 day input duration supplied to the GNN model. Map view of Kilauea along with the local seismicity during the eruption (c), and the first 0.75 days of radial GPS time series from station CRIM (d), with line darkness proportional to the cycle length. For the GPS time series in (b), the red curves are smoothed by a moving 30 minute Gaussian filter for each collapse cycle (the filter does not combine data from adjacent cycles).
• Figure 2: Comparison between predictions and observations for several different input window lengths. (a-c) GNN predictions for 0.5, 0.75, 1.0 day length inputs, respectively, for the GPS and tilt model. Predictions in (a-c) are average predictions from five repeated training runs of the model (with line widths marking $\pm$ 2 standard deviation). The first 29 events are used for training the GNN (blue), and the last 10 events are used for validation (orange). Results are compared with a 'null' model (d), which for each cycle is the average duration of all previously observed cycles (with line widths marking $\pm$ 1 standard deviation).
• Figure 3: Predictions of the timing of the final 20 collapse events throughout each cycle using the GPS and tilt model, including 10 cycles from training and validation. At each time step, subtracting the current time from the GNN cycle duration prediction results in the updated time to failure (blue circles). The true time to failure is shown by orange lines (solid for training, dashed for validation). Each prediction has line widths marking $\pm$ 2 standard deviation based on five training runs of the model.
• Figure 4: Predictions of the timing of the final 10 collapse events throughout each cycle (from validation) using different combinations of input data, including (a) seismicity, (b) all GPS stations, (c) GPS station UWEV, and (d) tilt. At each time step, subtracting the current time from the GNN cycle duration prediction results in the updated time to failure (blue circles). The true time to failure for these validation events is shown by dashed orange lines. Each prediction has line widths marking $\pm$ 2 standard deviation based on five training runs of the model.
• Figure 5: Predictions of the cycle duration of the final 10 collapse events (from validation) using different combinations of input data, including (a) GPS and tilt, (b) GPS, tilt and seismicity, (c) seismicity, (d) all GPS stations, (e) GPS station UWEV, and (f) tilt. At each time step, the absolute GNN cycle duration prediction values are shown (blue circles). The true cycle duration for these validation events is shown by dashed orange lines. Each prediction has line widths marking $\pm$ 1.5 standard deviation based on five training runs of the model.