Beneath California lies a labyrinth of faults—places where large chunks of rock slide past, over, under, or into one another. They move at the rate a fingernail grows until, eventually, they violently jolt and the ground shudders. Inevitably, destructive earthquakes will rock the state. Nothing can be done to prevent them from striking. But understanding where they are likely to happen, and what sort of power they can unleash, means communities can prepare.
In other words, the best way to mitigate these future disasters is to draw maps of the underworld, charting every fault and monitoring its behavior.
To create those maps, researchers deploy seismometers on the surface—little machines that sense and chronicle all sorts of vibrations, including earthquakes. Seismic waves are a bit like music. Identify their notes and rhythm changes, and scientists can work out what subterranean instruments created them, where they lie, and how they behave upon rupture.
This work, historically, has been painstaking, often slow, and sometimes imprecise. When Zachary Ross, an assistant geophysics professor at Caltech, started in the field in the early 2010s, he sought a new way forward. Traditional methods of quake-hunting science “just sucked, quite honestly,” he says. Even the best computer programs at the time missed certain earthquakes. There had to be a better way.
“We had massive amounts of data available,” says Ross. He explains that because California is so geologically active and so thoroughly covered in seismometers, there were more data than human experts alone could reasonably handle. Plus, most fault slips create tiny, imperceptible quakes. These emanate seismic waves so minuscule that even the most highly skilled seismologist can have difficulty spotting them in seismic data, especially when they resemble noise from human sources, like traffic.
In 2017 Ross had an epiphany. He saw machine learning programs handle huge sets of photos—identifying and categorizing elements within them, and with accuracy and speed that humans couldn’t match. So, he thought, why not apply a similar approach to seismology?
Ross’s first target: those copious tiny quakes. They might be harmless, but that doesn’t make them unimportant—their waves can illuminate each fault they pass through, including the more precarious, stress-loaded ones that may one day snap and trigger a disaster.
Ross and his colleagues took seismic waveforms from across Southern California that human scientists had identified as genuine quakes. Then he made templates of them, snapshots of each earthquake’s seismic wave pattern. Finally, he set an algorithm upon the seismic record, one searching for elusive quakes that matched those templates, those snapshots.
The algorithm quickly identified nearly two million previously hidden tiny quakes from 2008 to 2017—which, in turn, illuminated an intricate network of faults and fault features that prior quake searches were unable to see.
The results, published in 2019, “were so good that you had to even question if what you were seeing was really legitimate,” says Ross. “It’s such a cool paper,” agrees Marine Denolle, a geophysicist at the University of Washington who also uses machine learning in her research. “The body of work there is phenomenal.”
There was a drawback, however. This program, a sort of precursor to true AI software, could only find earthquakes in the seismic record that it was taught to recognize. Novel seismic events went unnoticed.
So Ross turned to more advanced tools: self-learning programs, software that could take preexisting information and make predictions about the future—in this case, what a vastly wider variety of earthquakes might sound like. Very quickly, these programs found all sorts of unfamiliar-sounding quakes—later verified by human scientists. “You just see so many things that were completely missed,” Ross says.
These machine learning programs are still evolving and have gone beyond identifying quiet quakes and hidden faults. They have been deployed all over California, where they have identified a new class of prolonged, slowly migrating earthquake swarms. In Hawaii they found a never before seen network of pulsing and migrating molten rock underneath two active volcanoes that traditional methods of seismic analysis were unable to identify.
“This is just light-years beyond what we could have done a few years ago,” Ross says. “It’s at superhuman levels now.”
Today many in the seismological community receive Ross’s work with cautious optimism. “I think it can really catapult the field of seismology forward,” says Wendy Bohon, a seismic hazards and earthquake scientist. AI multiplies and accelerates the capabilities of a single scientist. It can process many seismic records simultaneously, rendering them with great precision and in three dimensions faster than any human could in the same amount of time.
There is some concern that geoscientists who have not studied machine learning will have some catching up to do. “How can we train the broader seismology community to understand and know what goes on under the hood, so we can evaluate these products appropriately?” asks Bohon.
No comments:
Post a Comment