← Back to Sessions

Predicting Ground Motion Waveforms for Earthquake Early Warning Using Convolutional Long Expressive Memory Models

Description: 

Earthquake early warning systems are developed to mitigate immediate threats after the onset of large earthquakes and deployed worldwide including the west coast of the United states. Once detecting P-waves at sensors close to the epicenter, the system estimates the earthquake location, magnitude and fault geometry and predicts ground motion intensity parameters to issue a warning prior to the strong motion arrivals. Errors in the parameter estimation, especially magnitudes, lead to false alert or missing warning opportunities and the use of empirical ground motion models precludes accurate representation of the complex source, path and and site-term variabilities. Alternatively, ground motion intensities or waveforms can be forecasted by combining physics-based simulation and data assimilation, which can remove arrival detection and magnitude estimation, but their prediction accuracy remains insufficient.

We propose convolutional Long Expressive Memory Models, a novel sequence-to-sequence learning approach in machine learning (ML) to accelerate the inference time, while simulating complex spatio-temporal wave propagation phenomena. We test our approach by forecasting physics-based simulation waveforms for earthquakes along the Hayward Fault in the San Francisco Bay Area. We train the ML using >900 small point-source earthquakes (<0.5 Hz) recorded at sparsely distributed stations. Once trained, using the early portion of waveforms, the ML can predict the rest of waveforms on the fly at arbitrary sensor locations. The predicted and groundtruth waveforms show remarkable agreement, and the spatially heterogeneous peak ground velocities are well predicted. We then use the point-source trained model to predict large M finite-fault cases, a challenging domain shift problem for ML. Our ML method can predict up to M6 ground motions, suggesting strong potentials in generalization. To apply to real earthquakes, we envision to train ML models using many real small M earthquakes and a smaller number of larger M data, and achieve fast high-fidelity prediction based on observations.

Session: End-to-End Advancements in Earthquake Early Warning Systems - I

Type: Oral

Date: 5/3/2024

Presentation Time: 03:00 PM (local time)

Presenting Author: Nori

Student Presenter: No

Invited Presentation: 

Authors