Jeff J Mitchell / Getty Images / WIRED

The Holy Grail of seismology is earthquake prediction. Not forecasting, which gives a probability of a significant quake happening in an area within a period of years or decades, but actually predicting exactly when and where a major quake itself will strike, and what magnitude it will be.
Seismologists will unequivocally tell you that anyone claiming they can predict future earthquakes are false prophets. But that’s not to say that earthquake prediction is impossible. Scientists have invested a tremendous amount in achieving this hallowed goal, it’s just that so far they’ve come up short.

Advertisement

In December 2018, a coalition of researchers decided to try something new. They announced an online competition, open to anyone, in which participants had to predict future earthquakes being generated by a vice-like device in a laboratory. The twist? They had to design their own rudimentary artificial intelligences to make the predictions.
Thousands of people from across the world threw their hats into the ring. As reported by an article published in the Proceedings of the National Academy of Sciences earlier this month, the winning teams managed to come up with collections of code that managed to predict the timing of future laboratory earthquakes with striking precision.
It still isn’t clear how applicable this is to a real-life fault zone. But the promise of these new machine learning models implies that earthquake prediction isn’t a pipe dream, but a plausible possibility. And seeing as none of the victors had a background in seismology, this competition shows the benefits of casting an extremely wide net to find otherwise hidden talent – the sort that may one day save millions of lives.
Earthquakes have been around forever, but seismology is still in its youth. The sort of quantitative data we’re all used to today – magnitudes, shaking intensities and all – hasn’t been around for very long. The dawn of instrumental seismology only happened in the twentieth century. Ancient texts and indigenous knowledge describing earthquakes certainly help out fill in historical data deficits, and certain countries, like Japan, have a much longer written record talking about tremors and temblors.

Advertisement

But to study earthquakes properly, you need to document them as they happen. Short of inventing time travel, seismologists – at least compared to other geoscientists – will remain somewhat data starved.
“Our observational record in seismology is rather short,” says Kasey Aderhold, an earthquake seismologist at IRIS, a consortium of seismological researchers. That means that, despite huge advances in the past century, our understanding of the physics driving earthquakes are a little fuzzy and fairly theoretical.
There is, however, just enough data to peer into the future – to an extent. Organisations like the U.S. Geological Survey, using knowledge of past seismicity and present-day geophysical information, to say that, for example, there is a 20 per cent chance the San Francisco Bay area will experience a magnitude-7.5 quake in the next 30 years.
They can also forecast aftershocks, the quakes following on from the mainshock, the most powerful quake in the sequence. When a large earthquake (potentially a mainshock) happens, a series of equations and calculations – those derived from statistical models and reasonable assumptions about the way earthquakes work – churn out a time window (say, a week) in which a rough number of aftershocks with an approximate magnitude are likely to occur.

Advertisement

These calculations are very robust. But they are retrospective and reactive, and can’t be used to forecast mainshocks, the stars of the shows that can throw humans and their houses about. The problem is a lack of precursors: seismologists haven’t yet conclusively identified signals preceding mainshocks that clearly signpost their appearance and magnitude.
When it comes to earthquake prediction, “there’s been a long history of […] let’s call it stagnation, on this topic with very little serious success,” says Zachary Ross, a seismologist at the California Institute of Technology who wasn’t involved with the work. “Compared with the kind of progress that the climate people have made over the past several decades, which is just astounding, the existing techniques for [earthquake] forecasting seem to be fairly saturated in what they seem to be able to do.”
Enter, machine learning. Crudely put, this describes the ability of a computer code to absorb data, identify patterns, make choices or predictions, then learn from its mistakes to correct itself – all without significant human intervention. For seismologists, it’s a novelty; they have only begun discussing its potential and their work with it at major scientific conferences in the past few years.
And yet it’s already being used in the real world. A project that Ross was part of used machine learning to find millions of earthquakes buried in the seismological records of southern California. After being exposed to reams of seismic data, their software was able to quickly distinguish between random rumbles and the genuine grumbled of earthquakes, the sort imperceivable by humans.
Paul Johnson, a geophysicist at Los Alamos National Laboratory in New Mexico and lead author of the new study, thought that machine learning may be able to help with earthquake prediction. Instead of using equations designed around a human understanding of seismology, these codes would be starting afresh, consuming data and using that data alone to make predictions – and removing potentially erroneous or human assumptions from the mix.
An earlier study made use of an artificial quake-making machine in a laboratory. Steel blocks sandwiched a block of fault gouge, a rock typically found in natural faults. The blocks were mechanically moved around, pushing, squashing and pulling at the block. If the block cracks and there is a jolt forwards, voilà, you just made an earthquake.
Johnson and his colleagues wondered if a machine learning model, one provided with a stream of data on dimensions and characteristics of the ‘fault’, the stress and strain the system was undergoing and its resulting seismicity, could predict future earthquakes. “Very rapidly, a model was developed by some talented young people,” he says – a team of material scientists and mathematicians. Within months of setting the apparatus up, their software was able to predict with great accuracy when future quakes would transpire.
“It was just a revelation,” Johnson says. “It seemed like magic at first.” The model had managed to identify a handful of energy signatures in the seismic data that let it know the fault was about to fail.
But if this was a video game, it was set to ‘easy’ difficulty. The quakes were periodic, meaning they happened fairly regularly. Could machine learning handle irregular earthquakes, the sort you see in the wild?
Back in the 1800s, scientific competitions between competing researchers were commonplace, where white guys who often came from money threw shade at each other while trying to practically demonstrate the reality-explaining prowess of their theorems. In recent years, scientific competitions have become more focused on deriving technological solutions to problems, and normally involve team efforts.
Johnson and his colleagues didn’t see why this couldn’t apply to machine learning. They turned to Kaggle, a platform used by machine learning advocates to share research. In the past, it has been used to host competitions, including one where people tried to detect dark matter. In late-2018, they threw down the gauntlet: their laboratory fault machine was going to simulate earthquakes for the first half of 2019, and they wanted people’s bespoke machine learning models to predict when they were going to occur. The top five ranked teams would share $50,000 (£36,000).
Ultimately, 4,521 teams signed up, featuring 5,450 individuals. They came from a dizzying array of backgrounds, from mobile gaming to cartoonists, from insurance salespeople to those studying electrical signals produced by hearts and brains. Teams first built their models using training data: seismic signals, timing of the earthquakes, the shear stress the fault was experiencing, and so on. They then submitted up to two models per day to the adjudicators, who put them through their paces trying to predict the apparatus’ earthquakes while receiving only seismic data. Scores were given based on how precise these predictions were.
The top five teams – GloryorDeath, Reza, Character Ranking, JunKoda and The Zoo – became masters of quake prediction. The first placers, The Zoo, a team of eight members from the U.S. and Europe, were a mixture of acquaintances and complete strangers. Despite the potential for disorganised chaos, they managed to claim first place thanks to some clever hacks.
The first was to build their model not just using the training data, but the test data too, thereby making the examination an educational experience rather than just an exhausting gauntlet. In some circles, says Johnson, that would be considered cheating. But this is how machine learning would work in reality: it would learn from training data sets and from its experiences with genuine earthquakes.
The second winning feature was as counterintuitive as it was inspired: they fed noise into the data stream. Noise – caused by traffic, the wind, the ocean, animals or people walking about – is anathema to seismologists, who need to filter it out to hear earthquakes. It isn’t actually clear why this made their model more precise. “Some of the things people do just work, and you don’t necessarily understand why,” says Johnson. One possibility is that you are simply giving the models more data to train on and learn from. Practice makes perfect, after all.
Remarkably, none of the winners had a background in seismology. Were the adjudicators bruised by that revelation? Not at all, says Laura Pyrak-Nolte, an astronomer and physicist at Purdue University and study co-author. “For us, it was a tremendously exciting experiment.” And that real progress was made within a highly cooperative framework will only help to consign the irritating myth of the lone genius to history, says Aderhold.
Machine learning has already shown predictive capabilities in the Pacific Northwest’s Cascadia Subduction Zone. After listening to 12 years of the seismic soundtrack emanating from a very gradual fault movement named ‘slow slip’, it was able to find patterns in the noisy parts that predicated the next slow slip, a bit like knowing when the beat was about to drop in a song.
The next step is to try it out on relatively quiet faults that will one day violently fail and shake the land. “We’re really in the thick of that right now, and we just don’t know what the outcome will be,” says Johnson.
Laboratory quakes are simplified versions of bona fide faults. Success in this competition, then, doesn’t mean that machine learning has uncovered the Holy Grail. But it’s clear that cracking the case won’t be simple, if success is even possible. “It remains to be seen whether there will be advances in our ability to forecast real seismicity using machine learning,” says Ross.
But not knowing if there is a way to predict earthquakes won’t stop seismologists, and their newfound colleagues across the world, from trying. Aderhold brings up a Douglas Adams quote from his magnum opus, The Hitchhiker’s Guide to the Galaxy: “There is an art to flying, or rather a knack. The knack lies in learning how to throw yourself at the ground and miss.” Seismologists have hit the ground plenty in their quest, but they still dream of flight.
More great stories from WIRED
🌌 A rebel physicist has an elegant solution to a quantum mystery
🍪 Google is rewriting the web. Here’s the impact Chrome’s plan to kill cookies will have
😷 As more Covid-19 variants emerge, attention has turned to N95 and FFP2 face masks

Advertisement

🔊 Listen to The WIRED Podcast, the week in science, technology and culture, delivered every Friday
👉 Follow WIRED on Twitter, Instagram, Facebook and LinkedIn

Get WIRED Daily, your no-nonsense briefing on all the biggest stories in technology, business and science. In your inbox every weekday at 12pm UK time.

by entering your email address, you agree to our privacy policy

Thank You. You have successfully subscribed to our newsletter. You will hear from us shortly.
Sorry, you have entered an invalid email. Please refresh and try again.