Earthquakes are scary enough on their own, but their aftershocks sometimes cause even more damage. Now, researchers have developed algorithms that, they claim, provide more accurate predictions for those aftershocks.
Researchers have long tried to use physics to model the stresses on tectonic plates, so as to predict where earthquakes and aftershocks are likeliest. The crust may seem solid to us, but on long enough timeframes and under enough pressure, it behaves more like a fluid, and fluids have very complicated physics.
Since powerful new computers can now examine massive historical datasets of earthquakes, then-graduate student Phoebe Robinson DeVries wondered if machine learning might be able to speed up physics-based predictions. A paper she and colleagues published in 2017 showed that it sped them up by a factor of 50,000%. That’s right, four zeros.
Then they began working on whether the machine-learning-boosted models might be able to better pin down the locations of certain types of aftershocks. They trained their software on data from 130,000 real-world earthquakes and aftershocks.
The researchers tested their AI’s predictions using data from another 30,000 real-world earthquakes. Not only was the AI able to beat the existing standard for predicting the location of aftershocks, but DeVries and her colleagues were able to figure out which features of the earth’s crust it was using to improve its predictions, they report this week in Nature.
Bringing AI to the fight “is a really nice, efficient approach,” seismologist Lucy Jones of Caltech told Science News.
The study limited itself to one type of aftershock: the type that are caused by static tension that built up in new locations during the main earthquake. But some aftershocks are caused by so-called dynamic stress changes soon after the main event, and most studies, including this one, don’t yet try to account for those.
AI predictions are only as good as the data used to train the AI. In this case, the researchers used a global database of earthquake data, including something called slip, which measures how much the earth moved and where. But researchers often differ in how they map and record slip, so geophysicist Gregory Beroza, who was not a part of this study, sounded a note of caution in an accompanying piece in Nature.
Including this kind of software prediction into civil defense systems would also require it to run even faster. Right now it doesn’t approach real-time operation.
“Regardless of the physical interpretation, the performance of DeVries and colleagues’ artificial neural network is motivating,” Beroza wrote.