Five years ago, Paul Johnson wouldn’t have thought predicting earthquakes would ever be possible. Now, he isn’t so certain.
“I can’t say we will, but I’m much more hopeful we’re going to make a lot of progress within decades,” the Los Alamos National Laboratory seismologist says. “I’m more hopeful now than I’ve ever been.”
The main reason for that new hope is a technology Johnson started looking into about four years ago: machine learning. Many of the sounds and small movements along tectonic fault lines where earthquakes occur have long been thought to be meaningless. But machine learning—training computer algorithms to analyze large amounts of data to look for patterns or signals—suggests that some of the small seismic signals might matter after all.
Such computer models might even turn out to be key to unlocking the ability to predict earthquakes, a remote possibility that is so controversial, many seismologists refuse to even discuss it.
When the theory of plate tectonics gained ground in the 1960s, many scientists thought that earthquake prediction was only a matter of time. Once small quakes caused by shifting plates could be modeled, the thinking went, it should be possible to predict larger earthquakes days or even weeks in advance. But a multitude of factors, from rock type to the distance of a fault slip, affect the strength of an earthquake, and it quickly became apparent that models of small-scale tectonic activity couldn’t provide a reliable way to predict major earthquakes. Perhaps small shifts and slips, which occur hundreds of times per day, could indicate a slight increase in the probability of a large earthquake striking, but even after a swarm of minor tectonic activity, a big quake is still highly unlikely to occur. A better signal for an incoming earthquake is needed if prediction will ever become reality.
Using machine learning to find such a signal is likely a long way off—if it’s even possible. In a study published late last year, Johnson and his team suggested there could be a previously disregarded seismic signal that might contain a pattern revealing when a major earthquake—like the infamous and long-awaited Cascadia quake in the Pacific Northwest—could strike. If the hypothesis pans out, it could change the way earthquakes are forecast from seconds in advance to, maybe one day, decades in advance.
The most recent improvements in earthquake forecasting have been those precious seconds. Seismologists are working on improving early-warning systems like those in Japan and the ShakeAlert system being rolled out along the U.S. West Coast. Those systems send out alerts only after an earthquake has already started—but in time to shut down things like elevators or gas lines and warn communities farther from the epicenter.
Trying to extrapolate how big an in-progress quake is going to become, where its epicenter is and what’s going to be affected, all from a few seconds of data, is already a huge challenge, Johnson says. Existing warning systems have misjudged major earthquakes and given false alarms on others. But before 2007, we didn’t even have seconds’ notice. Where might we be in 2027?
“We don’t know how well seismology will really do a decade from now,” Johnson says. “But it will be much better than today.”
Advances in earthquake monitoring will likely rely on computers that have been trained to act as expert seismologists. With perfect memory, few preconceived notions and zero need for sleep, machines can sort through a sea of data collected as tectonic plates shift. All that information is comparable to what you would hear on a crowded street—the noises of cars, people, animals and weather all mixed together. Researchers sift through those signals, transcribed as waves, in an attempt to find out if any of them indicate an earthquake is happening or is about to happen. The hope has long been that, tucked into all that noise, there might be some sort of precursor that could be measured or observed to indicate the length of time until the next major quake.
One of those noises—what Johnson calls a “tremor-like signal”—has been identified and studied for a number of years. “I threw everything I had in my toolbox at it and decided there was nothing there,” he says.
But the algorithms and computers his team set up looked at the signal from a slightly different perspective, focusing on its energy. That energy (recorded as amplitude, a measure of the size of seismic waves) grew “ever so slightly” throughout the earthquake cycle, Johnson says. Once an earthquake hit, the signal’s amplitude dropped and restarted the cycle of regular growth until another quake hit.
It was a pattern.
That previously disregarded signal, Johnson says, “contained predictive information for predicting the next earthquake cycle” minutes in advance in the sped-up models of faults in the lab, which translates to decades in advance in real life. But results in the lab and the real world don’t always line up.
At this point, machine learning is not intended to help with earthquake prediction, but rather to understand quakes that have already started or quake dynamics in general. But advances in locating quakes, estimating magnitudes and sorting through “noise” all improve our understanding of how quakes act, including when they might strike.
“I want to make it clear that what we are doing is different than prediction. But, yes, all of these things are indirectly related,” says Mostafa Moustavi, a Stanford seismologist who is using machine learning to sort through background noise to detect small quakes.
Men-Andrin Meier, a seismologist at Caltech, says that his “best guess is that earthquakes are inherently unpredictable.” But nevertheless, he’s working on using machine learning to improve early warning systems, and improvements in the monitoring that goes into those alerts could potentially improve earthquake forecasts. Better maps of faults and better understanding of earthquake processes, trends and cycles could all go into improving forecasting, Moustafa says.
Even so, some seismologists think “prediction” is a fantasy. Robert Geller, a University of Tokyo seismologist, is well known for his pessimism about earthquake prediction.
“Earthquake prediction research isn't really a thing,” he says via email. “It just consists of gathering lots of data in the hope that a reliable ‘precursor’ can be found. None has ever been found to date.”
According to Geller, any lab results regarding earthquake signals can be ignored until they are reproduced consistently in the real world. “I have no doubt that they can find lots of apparent patterns in observed earthquake occurrence data looking backward. But I see no reason to think that such patterns will work going forward in time,” Geller says.
The Cascadia fault off Vancouver Island slowly slips all the time, producing low seismicity you can’t feel, and then lurches back into place about once a year. The very slight displacement of the Earth’s surface from that slipping can be monitored, so Johnson’s team tried to see whether the new signal their machine learning algorithms identified could predict the movement.
“And, lo and behold, it mapped to the displacement rate,” Johnson says.
The question now is how the signal might relate to the locking of the fault—the interlocked rocks that have kept the tectonic plates from slipping drastically and producing a major earthquake for about 300 years. Eventually, the locking of the fault will break, and a massive earthquake will strike. Perhaps the signal Johnson's team is studying, or another as-yet-undiscovered signal, could give some sense of when that will happen—if such signals are related to major earthquakes at all.