The shaking woke Thomas Heaton on a quiet winter morning in Pasadena, California. The streets were empty, with sunrise hours away. As Heaton lay in bed next to his wife, waves vibrated through their house. Ten seconds. Fifteen. Twenty. As a seismologist, Heaton had spent his career studying seismic waves like these. By feel and duration, he guessed this quake was big, maybe a magnitude 6.5, and close, under west Los Angeles. Plenty dangerous. An aftershock rolled through. He was needed. "I've got to get to work," he told his wife.
At the time, in 1994, Heaton was the lead scientist at the earthquake field office of the U.S. Geological Survey (USGS) in Pasadena. He drove to the office in darkness, imagining the fires, collapsed bridges, and crumbled buildings closer to the epicenter. At the office, seismic readings partially validated his gut: "I was right about the magnitude and approximate distance," he says—though not the location. The quake had struck farther north, under the neighborhood of Reseda, on a previously unknown fault. The Northridge quake, as it came to be known, killed 57 and caused many billions of dollars in damage. There had been no warning, no sirens sending people into the streets. Heaton recalled how he had guessed the size of the earthquake when the first, gentle waves reached his bedroom. There must be a way, he thought, to translate his gut check into a short but useful warning.
After decades of work, Heaton's dreams have taken form. Last month, USGS unveiled ShakeAlert, the West Coast's earthquake early warning system. If all goes as planned, a dense network of seismometers in California, Oregon, and Washington will detect the first, weak waves of an earthquake and relay a rapid warning of ground shaking to come. To start, those warnings will go to first responders, power companies, and transit agencies. But in the next couple of years, alerts could roll out to the public to provide at least a few seconds of warning. Not much time, but enough to "drop, cover, and hold on," says Doug Given, a geophysicist in Pasadena who is leading the USGS effort.
For years, ShakeAlert was an academic side project of California seismologists, especially the gravelly voiced Heaton, now at the California Institute of Technology (Caltech) in Pasadena, and Richard Allen, his soft-spoken counterpart at the University of California (UC), Berkeley. They were inspired by warning systems in Mexico, Japan, Taiwan, and Chile, among others, which emphasize detecting earthquakes at the source and warning distant cities before the seismic waves arrive. Many people thought such a system would be useless in fault-riddled California, where earthquakes seem to erupt underfoot anywhere. But Heaton and Allen persevered, deploying a pilot system in 2012.
Now, politicians are offering their support. Last year, some $13 million in annual funding flowed in from the federal government, along with $10 million more for sensor upgrades; California kicked in another $10 million. In his state-of-the-city speech last year, Los Angeles Mayor Eric Garcetti pledged: "By the end of 2018, we will deploy an earthquake early warning system to every corner of this city—in schools, at businesses, even on your smartphone."
This year's version doesn't quite measure up to that promise. Only half of the system's 1675 seismic stations have been installed. The technology to rapidly push alerts to mobile phones is not mature. And the public has yet to be trained in how to respond to such alerts, which are sure to include false alarms.
The system's scientific ambitions have also been humbled. The scientists developing ShakeAlert once promised it could warn of strong, violent shaking from a distant earthquake far in advance. That pitch stemmed especially from Allen, Heaton's friendly rival, who believed the final magnitude of an earthquake was determined by its first few seconds of rupture. If so, the system could catch an earthquake rupturing on a remote section of the San Andreas fault and give Los Angeles 1 minute or more of warning of severe shaking.
But over 15 years of development, reality has intruded: Faults fracture in complex, unpredictable ways. The current incarnation of ShakeAlert might offer 10 seconds of warning for a severe event—if you're lucky, Heaton says. "We're back to the simple ideas and just making the engineering part of this problem work," he says. "We're just trying to get it born."
Inspiration from Japan
The son of a mathematician at Rutgers University in New Brunswick, New Jersey, Heaton, born in 1951, grew up on the stable ground of the East Coast. Dyslexia, which made molecular structures a jumble, pushed him out of chemistry into physics, but he didn't find his calling in the Cold War tasks of the time, either. "They had enough nuclear weapons to blow up the entire solar system and they didn't need any more," he says. Instead he was drawn to study Earth's own convulsions. As a graduate student at Caltech he experienced his first earthquake, an aftershock of the San Fernando quake in 1971. During that disaster, emergency workers took 3 hours to figure out where the heavy damage was. Seismologists were little help.
Heaton had three children, so he took a job with Exxon. He lasted less than a year, but while there he learned that Japan was already using early earthquake warnings to shut down bullet trains. "At that point, I got very excited," Heaton says in his Caltech office, where gas mains buckled by earthquakes serve as table stands.
He laid out his idea for a U.S. system in a 1985 paper in Science. Because seismic waves travel far more slowly than electrical signals, a "seismic computerized alert network" could detect an earthquake at its source and relay a warning of ground shaking to cities far from the epicenter. Automated systems could act immediately to prevent chemical spills, electrical fires, and other catastrophes. Such a system would do little to protect San Francisco from an earthquake like the one in 1906, which was centered near the city. But it could give minutes of warning for great quakes that start far from populated regions. It was a simple model with many assumptions—including, critically, an immediate detection of an earthquake's magnitude. "We can do it in 10 years," Heaton promised anyone who asked.
It took longer. But as he climbed the ranks at USGS, Heaton updated Southern California's network of seismometers toward the always-connected compatibility needed for early warning. He also formed an alliance with Hiroo Kanamori, a decorated Caltech seismologist. Others in the field had spent years fruitlessly debating whether earthquakes can be predicted. Kanamori saw a better use for seismologists' talents: developing a warning system for earthquakes already underway. By the early 2000s, Allen, then an ambitious postdoc, had joined their effort.
Like Heaton, Allen was a transplant from stable terrain, namely, the United Kingdom. He, too, came to Caltech dissatisfied with sterile debates, in his case about Earth's internal structure. Early warning, it seemed, was the rare scientific discipline that could save lives. And at the time it appeared poised for a breakthrough.
You’re not going to get much time. If it’s going to be dangerous, we won’t know that till the last seconds.
At their most basic, earthquakes result when the strain built up between two locked chunks of Earth's crust becomes too much to bear and the slabs of rock slip past each other along a fault. The larger the slip area, the bigger the earthquake. As the rupture starts, it tosses off pressure (P) and shear (S) waves. P waves percuss the rock like a drumstick, traveling quickly through incompressible material. S waves, though more powerful, struggle through the rock because of their sashaying motion and lag well behind.
The classical view had been that nothing about the first waves from a rupture indicates how it will grow, reflecting an inherently chaotic, unpredictable system. But in the 1990s, lab-built models and some data on actual earthquakes suggested that a nucleation phase—a brief period of subtle slipping at the quake's start—could predict the size of the ensuing rupture. If that were true, forecasting the ultimate magnitude of an earthquake from only a few seconds of P waves might be possible. That ability could power a potent early warning system—a possibility that Yutaka Nakamura, an earthquake engineer at a private company in Japan, had already begun to pursue to improve bullet train warnings.
Allen and Kanamori built on Nakamura's work in a 2003 Science paper. In records from 53 California earthquakes, the largest a magnitude 7.3, they found a correlation between the time the initial P wave took to complete one cycle, called τ, and the resulting magnitude. That relationship became the core of an algorithm Allen developed called ElarmS. It led him to argue, in a 2005 Nature paper, that earthquakes are deterministic, their fate structured by their start, contrary to the conventional wisdom. "That paper," he notes, "was very controversial."
Heaton, though, doubted a chaotic system such as an earthquake would surrender its secrets to a simple equation. He recalled how his gut feeling and knowledge of past events had called out the Northridge quake. He started to develop code to re-create that intuition. As with ElarmS, the code relied on P waves from the first few seconds of a quake. But instead of using τ to leap to a final magnitude, the system compared the features of the initial waves with those of past quakes to create a digital gut check. Heaton called the project Virtual Seismologist.
Despite the ongoing debate, USGS began to finance Heaton, Allen, and other teams to work on the algorithms that make up the core of ShakeAlert.
A lesson from the Big One
When a magnitude-9.1 earthquake struck 70 kilometers off Japan on 11 March 2011, the country's warning system was little help for people in the path of the torrential tsunami that swamped the coast; nearly 16,000 died. But the system did alert more than 50 million people and halt bullet trains and elevators in many regions before the shaking began. It also served as a wake-up call for U.S. researchers to push for their own system. "That was the tipping point," Allen says.
At a 2-day emergency summit at UC Berkeley a month later, the ShakeAlert team won a $6.5 million commitment from the Gordon and Betty Moore Foundation in Palo Alto, California, to build a prototype. USGS was sold, too, and agreed to run the system. The funders accepted that ShakeAlert need not be perfect; the Japanese public had appreciated the Tohoku warning despite its flaws. "They just had to make sure it worked reasonably well," Heaton says. And that meant solving the problems that Tohoku had exposed.
Despite the warning's success, it failed to alert Tokyo residents, far south of the quake, who were blindsided when the ground began to shake. The problem was that the system had located the earthquake to a single point. It then calculated how the shaking at that point, the hypocenter, would affect more distant locations. For earthquakes of magnitude 6.5 or smaller, which rupture for only a few seconds, that approach is reasonable. But the Tohoku fault rupture grew toward Tokyo, extending to some 400 kilometers over more than 3 minutes. "One thing we did not expect is that really long fault rupture," says Masumi Yamada, a seismologist at Kyoto University in Japan who studied with Heaton. As a result, the alert underestimated the quake's magnitude and extent.
The algorithms developed for ShakeAlert had the same shortcoming. "We realized really quickly that if there was a major earthquake along the southern San Andreas fault, we wouldn't expect shaking in Los Angeles because it was so far away," says Maren Böse, a seismologist at ETH Zurich in Switzerland who had also worked with Heaton. But while Yamada was at Caltech, Heaton worked with her to develop a way to track the growth of a rupture in real time, by measuring the shaking along its path. Böse, Heaton, and others then refined that technique. Virtual Seismologist yielded to an algorithm called the Finite-Fault Rupture Detector (FinDer), which updates its warnings as an earthquake progresses. FinDer, despite its late start, proved vital to showing that ShakeAlert could handle a Tohoku-size strike. "And really, if we can't do big earthquakes," Heaton says, "we're missing the point."
Humbled by reality
A klaxon sounded eight times, followed by an insistent robotic voice: "Earthquake. Earthquake." Heaton, seated at his desk, had a map of Southern California on screen. A simulated earthquake had just struck at the southern end of the state, by the Salton Sea, with an estimated magnitude of 7. The quake posed little threat to Pasadena, 250 kilometers to the north, and so ShakeAlert warned of only light shaking. But the rupture didn't stop there, and FinDer stayed on the case. A gray line began to extend toward Los Angeles, as did expanding rings of yellow and red: the warning P waves and damaging S waves. "Now it's getting closer," Heaton said. "And bigger as it goes."
"Earthquake. Earthquake. Moderate shaking expected in 42 seconds," the voice warned. The estimated magnitude had gone up to 7.8—a 16-fold leap in energy. The rupture continued, and ShakeAlert upped its warning again: "Strong shaking expected in 23 seconds. Earthquake. Earthquake." Finally, 7 seconds before the damaging waves arrived, ShakeAlert gave its final warning: "Very strong shaking expected." The klaxon fired rapidly. And then silence. "You're not going to get much time," Heaton says. "If it's going to be dangerous, we won't know that till the last seconds."
Heaton says he still wishes that some signal buried in the first moments of an earthquake could reveal more. But even before Tohoku, the grand promise of predicting an earthquake's final magnitude from its first moments had begun to fall apart. In records for earthquakes with magnitudes above 7, "We started seeing a saturation effect," says Gilead Wurman, one of Allen's former students at UC Berkeley. "You'd start to underestimate the magnitude."
And really, if we can’t do big earthquakes, we’re missing the point.
Heaton's recent work, conducted especially with Men-Andrin Meier, a seismology fellow at Caltech, has only solidified doubts. A 2016 comparison of P waves recorded within 25 kilometers of the hypocenter of earthquakes in the United States, Japan, and elsewhere showed that the small and large quakes looked identical at the start. The determinism of the nucleation phase, it seemed, was a ghost.
Meier and Heaton, along with Pablo Ampuero, another Caltech seismologist, have found that as an earthquake develops, it does drop a hint about its ultimate strength. In a database of 116 earthquakes greater than magnitude 7 created by a former postdoc, Lingling Ye, now at Sun Yatsen University in Guangzhou, China, they found that once the rupture starts to slow, the median earthquake ends up no more than doubling in strength. "At some point you see it slowing down, and then you know after that it's all downhill," Ampuero says. Heaton believes that effect, which they call weak rupture predictability, is the only pattern they'll be able to tease out. But it emerges late and has little predictive value for individual earthquakes. There's no sign of a clear connecting thread from the start to the end of an earthquake.
Faced with the mounting evidence that determinism isn't holding up, Allen, too, is settling for weak predictability, something his own recent work has supported. That happens in science: Careers are made by staking out either side of a data-poor claim, and then a middle ground emerges. Yet even though most seismologists now agree that the start of an earthquake does not determine its end, many still think its early stages might somehow influence whether the rupture can grow by jumping faults or sections of locked rock. The earthquake's start may not drive all action, but it may still be a prologue that—in some way still not evident in the data—informs the rest of its story.
A real-world test
In September, while Allen was riding a Bay Area Rapid Transit train near San Francisco, his rail car ground to a halt. The conductor's voice came on the intercom. A magnitude-3.3 quake had struck 40 kilometers north of Berkeley, and the train system, following protocol, had stopped for safety. "I can't believe it: We have seen Yosemite, San Francisco, and now we have been in an earthquake!" one family of tourists said. The unplanned stop delighted Allen, too. "After working on this for over a decade, here it was in action and I was on the receiving end."
Over the past year, Given has pushed the ShakeAlert team to meld its unruly competing algorithms into a cohesive whole. First, a fast-estimating code called EPIC—consisting primarily of Allen's ElarmS—generates an initial magnitude, treating the quake as a point source. But if EPIC sees a quake lasting more than a few seconds—and therefore larger than magnitude 6.5—FinDer takes the lead, tracking the rupture from there and updating the magnitude. The refinements will continue. "It's been kind of a closed club through this year," Given says, but the agency is now soliciting other researchers to improve the code.
New technologies will sharpen the warnings, too. GPS sensors, though slower than seismometers, can capture even shaking strong enough to max out conventional instruments, enabling the system to cope better with the biggest earthquakes. And Heaton expects artificial intelligence, especially neural networks, will in the next few years be able to discern P waves, an earthquake's first whisper, from seismic noise earlier than the existing algorithms. At first, Heaton was skeptical of the technology. "But then it dawned on me that this other neural network was in many ways more capable than this neural network," he says, pointing at his head.
After working on this for over a decade, here it was in action and I was on the receiving end.
Any warning system is only as good as its messaging, and how ShakeAlert will best reach the general public remains uncertain. "The technology for doing rapid massive alerting doesn't exist in the United States," Given says. The cellular messaging system that handles child abduction or severe weather alerts wasn't designed to relay warnings in seconds—more like minutes. Los Angeles will begin to test an alternative, using notifications on a smartphone app, but the fear is that such a system could easily overload.
USGS has set one important parameter: Instead of waiting until a risk is severe, ShakeAlert will skew toward more alerts, sounding an alarm once a location is at risk of "light shaking." That will increase the warning time—but it also will mean that, if the rupture grows, the prediction could change to severe shaking only seconds before hitting. And the public might grow complacent about those alarms and fail to respond to the rare mild threat that, in a moment, turns severe.
The faults riddling Heaton's adopted state guarantee that soon, ShakeAlert will get its first high-profile test. "It's a little terrifying," he says. "The world will be watching. Here's your chance to sing in front of everybody. You just hope you don't—" And Heaton's gravelly voice broke into a croak that echoed down the hall.