Predicting earthquakes with AI

Posted: July 30, 2024

Fifteen earthquakes rocked the Sichuan and Yunnan regions of China over seven months in 2021. A new AI created by teams at UT Austin predicted fourteen of them a week in advance—though it also predicted eight others that never happened. The AI used 5 years of seismic recordings from sensors throughout the region to pick up on patterns of tremors that tended to precede earthquakes.

Just a decade ago, professional seismologists considered earthquake prediction to be tantamount to astrology. But these new AI predictions in China constitute just one of many recent AI earthquake prediction successes that now have scientists hunting in earnest for increasingly reliable prediction tools.

Instead of training AI on historical data, like the Austin team, Paul Johnson, a Senior Scientist and Fellow at Los Alamos National Laboratory in New Mexico, trains AI on sensor data from artificial earthquakes, which he creates by putting pieces of rock under pressure in the lab. Johnson discovered that low-amplitude acoustic signals from the rock—which scientists had previously dismissed as mere noise—allowed the AI to predict when fault lines in the rock would undergo a quake.

Using the same AI method to analyze data from seismic sensors on Vancouver Island, his team identified a similar acoustic signal that predicts the rate at which the Cascadia fault in the North American Pacific Northwest is moving. He can’t predict when there will be an earthquake along the fault (like those that rocked Vancouver Island earlier in July), but the finding gives scientists one more piece of data to work with.

“[Machine learning] allows you to make these correlations you didn’t know existed. And in fact, some of them are remarkably surprising,” Johnson told Dr. Allie Hutchinson for the MIT Technology Review.

AI can’t yet predict earthquakes with anything like the accuracy of weather forecasts (which have improved dramatically over the last 40 years so that they now have about 90% accuracy 5 days out—largely thanks to bigger, more accurate sensor data). Several locales do have sensor arrays employing algorithms that will give them a 30- to 60-second warning to brace for an incoming quake once it has started—precious seconds in which to take shelter or turn off systems like natural gas and electric. But the holy grail is long-term forecasts, something AI has made more progress on in the last decade than most had previously dared hope.