核心概念
Deep learning graph neural networks can accurately predict the timing of caldera collapse events at Kīlauea volcano using only a fraction of the available monitoring data.
要約
The content describes the application of deep learning methods, specifically graph neural networks (GNNs), to forecast the timing of caldera collapse events at Kīlauea volcano during the 2018 eruption. Key highlights:
- The 2018 Kīlauea eruption featured over 60 quasi-periodic caldera collapse events, providing a unique dataset to test forecasting methods.
- The authors trained a GNN to predict the time-to-failure of the collapse events using a combination of GPS, tilt, and seismicity data recorded at the volcano.
- The GNN model was able to predict the collapse times with high accuracy, outperforming non-machine learning based forecasting methods.
- Predictions improved with increasing input data length, and were most accurate when using high-SNR tilt data.
- Applying the trained GNN to synthetic data revealed that the model was sensing the underlying physics of the caldera collapse process.
- The results demonstrate the potential of machine learning methods for forecasting real-world catastrophic events with limited training data.
統計
"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."
"For input lengths of 0.5 and 1.0 days, the combined GPS and tilt model obtains average validation residuals of 3.89 hours and 1.75 hours, respectively."
"The null model predictions are significantly less accurate compared to the GNN predictions. For example, for 0.5 day length inputs, the RMS residual of the null model on the validation data is 14.1 hours, compared to 3.9 hours with the GNN."
引用
"Forecasting the timing of catastrophic geophysical events such as volcanic eruptions or earthquakes is a long standing challenge in geophysics."
"Our findings reveal that for this well monitored sequence, cycle durations were (retroactively) predictable to within a few hours given data from only a fraction of the mean recurrence interval."
"The predictability of these large scale collapse events demonstrates the potential of ML-based forecasting of significant, real-world hazards."