In Japan, the new year began with disaster as a 7.5 magnitude earthquake struck the Noto peninsula on the country’s western edge on Monday. Dozens more aftershocks, many measuring between four and six in magnitude, shook near the coastal epicenter in the hours since, and scientists warn that more are expected in the coming days. The Japan Meteorological Agency issued a major tsunami warning for much of Ishikawa prefecture, where the quake occurred, but has since reduced it to an advisory. As of Tuesday afternoon, at least 55 people are confirmed to have died.
More than 2,000 active fault lines lie beneath Japan, making it one of the most earthquake-prone countries in the world, with a long legacy of tragic seismic events. As a result, like many other at-risk places, Japan has invested heavily in disaster-proof buildings and earthquake warning systems over the last few decades.
But at the center of earthquake preparedness is, arguably, the field’s most controversial area of research: predicting when a quake will strike. Many scientists have long considered earthquake forecasting to be impossible—or, at best, they have approached it with extremely tempered optimism.
Even as recently as 2013, “the very topic of earthquake prediction was deemed unserious, as outside the realm of mainstream research as the hunt for the Loch Ness Monster,” writes seismologist Allie Hutchison for MIT Technology Review. And the U.S. Geological Survey maintains that “neither the USGS nor any other scientists have ever predicted a major earthquake.”
But given recent improvements in artificial intelligence, some researchers have been studying whether that could change. “I can’t say we will, but I’m much more hopeful we’re going to make a lot of progress within decades,” Paul Johnson, a seismologist working with machine learning at Los Alamos National Laboratory, told Smithsonian magazine’s Matthew Berger in 2019. “I’m more hopeful now than I’ve ever been.”
Last fall, researchers at the University of Texas at Austin bolstered such hopes for earthquake prediction with a seven-month trial in China. In their study, published in the Bulletin of the Seismological Society of America in September, an A.I. algorithm correctly predicted 70 percent of earthquakes one week before they happened. The team trained the A.I. on five years of seismic recordings, then asked it to locate upcoming quakes based on current seismic activity.
In all, the algorithm successfully forecasted 14 earthquakes, each within about 200 miles of its actual epicenter. Meanwhile, it missed one quake and predicted eight that never happened.
The trial was part of an international A.I.-design competition, one of a few such events held in recent years to advance earthquake prediction technologies.
“Predicting earthquakes is the holy grail,” Sergey Fomel, a geoscientist at UT Austin and a member of the research team, says in a statement. “We’re not yet close to making predictions for anywhere in the world, but what we achieved tells us that what we thought was an impossible problem is solvable in principle.”
Additionally, machine learning could help seismologists detect hidden patterns in data or collect more data to better inform earthquake forecasting, Hutchison writes for MIT Technology Review. For example, some researchers are showing how A.I. might use recordings from a specific seismic site to anticipate an earthquake’s magnitude. One team has built and trained neural networks to predict where aftershocks may occur after an initial strike. And others are using machine learning to identify and extract seismic waves—the vibrations that propagate through the earth during tectonic activity—from a noisy record of other rumblings in the ground.
According to the World Meteorological Organization’s Focus Group on A.I. for Natural Disaster Management, earthquakes accounted for 21.8 percent of all natural disasters to have A.I. models applied to them for the purpose of risk reduction between 2018 and 2021. And the growing body of research on earthquake prediction is raising scientists’ hopes that more breakthroughs may be possible.
“You know, there’s tremendous skepticism in our community, with good reason,” Johnson tells MIT Technology Review. “But I think this is allowing us to see and analyze data and realize what those data contain in ways we never could have imagined.”