Researchers reveal at the Seismological Society of America (SSA) 2021 Annual Meeting that a deep spatiotemporal neural network trained on more than 36,000 earthquakes offers a new method of swiftly predicting ground shaking strength once an earthquake is underway.
Based on the characteristics of the first seismic waves to be recorded following an earthquake, DeepShake analyzes seismic signals in real time and sends out advanced warnings of heavy shaking. It was created at Stanford University by Daniel J. Wu, Avoy Datta, Weiqiang Zhu, and William Ellsworth.
Seismic data from the 2019 Ridgecrest, California sequence was used to train the DeepShake network.
The neural network sent simulated alerts between 7 and 13 seconds before the arrival of high intensity ground shaking to places in the Ridgecrest area when its creators evaluated DeepShake’s capabilities using the actual shaking of the July 5, 2011, magnitude 7.1 Ridgecrest earthquake.
The novelty of utilizing deep learning for quick early warning and forecasting based solely on seismic recordings was emphasized by the authors.
“DeepShake is able to pick up signals in seismic waveforms across dimensions of space and time,” explained Datta.
DeepShake demonstrates the potential of machine learning models to improve the speed and accuracy of earthquake alert systems, he added.
We’ve noticed from building other neural networks for use in seismology that they can learn all sorts of interesting things, and so they might not need the epicenter and magnitude of the earthquake to make a good forecast. DeepShake is trained on a preselected network of seismic stations, so that the local characteristics of those stations become part of the training data. When training a machine learning model end to end, we really think that these models are able to leverage this additional information to improve accuracy.
Daniel J. Wu
“DeepShake aims to improve on earthquake early warnings by making its shaking estimates directly from ground motion observations, cutting out some of the intermediate steps used by more traditional warning systems,” said Wu.
According to Wu, many early warning systems first assess the position and magnitude of an earthquake before calculating the ground motion for a specific place using ground motion prediction equations.
“Each of these steps can introduce error that can degrade the ground shaking forecast,” he added.
The DeepShake team used a neural network strategy to address this. A neural network’s collection of algorithms is trained without the researcher specifying which signals are “essential” for the network to consider when making predictions. Directly from the data, the network learns which features best predict the intensity of upcoming shaking.
“We’ve noticed from building other neural networks for use in seismology that they can learn all sorts of interesting things, and so they might not need the epicenter and magnitude of the earthquake to make a good forecast,” said Wu.
“DeepShake is trained on a preselected network of seismic stations, so that the local characteristics of those stations become part of the training data. When training a machine learning model end to end, we really think that these models are able to leverage this additional information to improve accuracy,” he said.
Wu, Datta and their colleagues see DeepShake as complementary to California’s operational ShakeAlert, adding to the toolbox of earthquake early warning systems.
“We’re really excited about expanding DeepShake beyond Ridgecrest, and fortifying our work for the real world, including fail-cases such as downed stations and high network latency,” added Datta.