Deepshake: Earthquake Early Warning with a Deep Generative Spatiotemporal Recursive Neural Network

Session: Earthquake Early Warning: Current Status and Latest Innovations
Type: Oral
Date: 4/29/2020
Time: 09:15 AM
Room: 115
Description: 

With the recent public activation of the system ShakeAlert system in California, earthquake early warning is a highly applied area of research. Provided a set of ground motion measurements from a network of seismic monitoring stations in real-time, early warning systems aim to predict the intensity of earthquake ground motion at locations both near and far before strong shaking arrives.  In this work, we construct DeepShake, a deep generative spatiotemporal recursive neural network to predict shaking intensity in space and time. DeepShake is trained on ground acceleration using a dataset of 35,679 earthquakes from the USGS earthquake catalogue that occurred between June 1^st, 2019 and September 30^th, 2019 as recorded by 15 stations within 120km of the Ridgecrest, California, sequence. We convert 3-component accelerograms into their high-pass (1 Hz) envelope, convert the triaxial acceleration into its absolute magnitude and then downsample by calculating a one second running average over the signal. The network was not given any a priori knowledge of station locations and learns wave propagation amplitudes and delays solely from training data. DeepShake is a network forecasting model, able to predict shaking 10 seconds in the future at all stations given shaking observed at only a single station. Averaged over 7,136 validation earthquakes, DeepShake correctly predicts the ground acceleration in each 1-second window with a mean error of 6% and achieves a 8.9% equal error rate for MMI alert thresholds. The model creates an alert within 2 seconds from the first wave arrival on the network with an accurate amplitude forecast across the network for both principal earthquakes, M_W 6.4 and 7.1, of the Ridgecrest sequence.

Presenting Author: Daniel J. Wu

Authors