An earthquake early warning system that uses artificial intelligence (AI) to predict how the ground will move during a temblor can give several seconds' advance notice that the shaking is coming.
A similar system that uses more traditional computing power already exists on the U.S. West Coast. It's called ShakeAlert, and it works by detecting the first waves of earthquake motion — called P waves — and then calculating when the set of waves that cause most of the shaking — slower-moving S waves — will arrive.
The new system in development is called DeepShake, and it is also intended to provide a few seconds' warning of imminent shaking once an earthquake has started. However, DeepShake uses a deep neural network, a type of AI learning, to identify patterns from past earthquakes in order to predict how the shaking from a new quake will travel. This could lead to faster processing and easier generalizability across different earthquake-prone regions.
"When we set out on this project our goal was to beat the ground motion prediction equations that are currently used" to program shake-alert systems, said Avoy Datta, a master's student in electrical engineering at Stanford University who was part of the team that developed DeepShake. "They tend to be very slow. You need numerical solvers, running on supercomputers, and they can take minutes and hours to process."
In contrast, "If we run 25 DeepShake models, it takes around 6.1 milliseconds on a single research GPU [graphics processing unit]. " Datta told Live Science. “This is going to be blazing fast.”
In a presentation on April 23 at the virtual meeting of the Seismological Society of America, Datta and his colleague, Stanford undergraduate Daniel Wu, reported their results after training DeepShake to predict the ground motion of earthquakes near Ridgecrest, California. Ridgecrest is in the seismically active Eastern California Shear Zone, and in 2019, a sequence of earthquakes shook the region. The largest, a magnitude 7.1 quake, hit on July 5.
Datta, Wu and their colleagues used this earthquake sequence to train DeepShake to predict ground shaking in the area. They started with a dataset of more than 36,000 quakes that struck Ridgecrest from July to September 2019 (most were quite tiny). They fed 80% of the dataset into the deep neural network, saving 10% for tweaking the parameters of the network and a final 10% for testing whether the network's outcomes matched reality.
The researchers programmed the network to assign more weight to the larger earthquakes in the sequence, which were relatively few, so that it could perform better as an early warning system — after all, the largest quakes are the ones that people need warning about the most.
Despite the fact that DeepShake was given no information about the earthquake's location or type, it was able to warn of shaking at other seismic stations in the network between 3 and 13 seconds before it happened, Wu told Live Science. This is similar to the amount of advanced notice with ShakeAlert. Wu and Datta don't view this other system as a competitor, however. Rather, they said, DeepShake technology could be used to complement ShakeAlert. The researchers hope to expand the testing to other faults and earthquake sequences.
Ground shaking at any given spot can be tricky to predict. For example, ShakeAlert failed to send out warnings during the largest quakes in the Ridgecrest sequence in 2019 because the shaking was expected to not reach the program's threshold of "light shaking" in some areas that did indeed experience light shaking. The developers of ShakeAlert have made changes to it since 2019 to incorporate those lessons learned. The advantage of deep-learning networks, though, is that they automatically incorporate quirks of the site, because they are based on past experiences of shaking in that location, Wu said. Unlike ShakeAlert, which uses more universal equations with assumptions built in, DeepShake would have to be retrained in each individual region where it was used. This training, however, would catch patterns that traditional equations might not.
"The places where deep learning really flourishes are places where there is lots of lots of data and lots of complicated patterns to uncover," Wu said.
Originally published on Live Science.
Live Science newsletter
Stay up to date on the latest science news by signing up for our Essentials newsletter.
Stephanie Pappas is a contributing writer for Live Science, covering topics ranging from geoscience to archaeology to the human brain and behavior. She was previously a senior writer for Live Science but is now a freelancer based in Denver, Colorado, and regularly contributes to Scientific American and The Monitor, the monthly magazine of the American Psychological Association. Stephanie received a bachelor's degree in psychology from the University of South Carolina and a graduate certificate in science communication from the University of California, Santa Cruz.