Oklahoma never used to be known for its earthquakes. Before 2009, the state had roughly two quakes of magnitude three and above each year. (Magnitude three is when things shake on the shelf, but before houses start getting damaged.) In 2015, this tally rocketed to more than 900, though it’s calmed since, falling to 304 last year.

This sudden increase is thought to be caused by the disposal of wastewater by the state’s booming fracking industry, and it’s caught seismologists off-guard. As a historically quake-free area, Oklahoma doesn’t have enough equipment to detect and locate all of these quakes, making it hard to investigate their root cause. “There are no major faults in Oklahoma so it’s just not something we would expect,” Thibaut Perol, a deep learning researcher who’s worked on this problem, tells The Verge. “And to understand what’s happening, we need a big, big catalogue of earthquakes.”

The solution proposed by Perol and his colleagues from Harvard University’s engineering and earth sciences departments is to use artificial intelligence to amplify the sensitivity of the state’s earthquake detectors, otherwise known as seismographs. In a paper published today in the journal Science Advances, they show how effective this technique is — capable of detecting 17 times more earthquakes than older methods in a fraction of the time.

The method is similar to the voice detection software used by digital assistants like Alexa and Siri, explains Perol. It’s all about uncovering the signal hidden in the noise. With Alexa, that means listening out for your voice commands while ignoring the background sound of your home. And for seismographs, it means cancelling out the normal geological rumblings of the Earth (what’s known as “ambient seismic noise”) to spot the earthquakes that might be very small or far away. This way, scientists in Oklahoma can get more out of the data they already have.

To make this happen, Perol and his colleagues trained a convolutional neural network to recognize background noise, feeding it data from seismically quiet areas, like pre-fracking era Oklahoma and the geological dead-zone of Wisconsin. (The state has only really had one significant earthquake, and that was in 1947.) As with all neural networks, the software examines this input and learns to pick out common patterns. Once it knows what ambient rumblings sound like, it can remove these from the data, leaving behind the tiny earthquakes that had previously been hidden — like sea shells revealed by a retreating tide. As a bonus, the neural network is even able to identify the rough whereabouts of individual quakes by matching the patterns they created with historical data where a tremor’s location was known.

“With this method we are able to detect earthquakes of magnitude zero or minus one, and these are signals you wouldn’t be able to see with a human eye,” says Perol.

William Yeck, a seismologist at the United States Geological Survey (USGS), praised the work as “compelling and novel.” Speaking to The Verge by email, he noted that the neural network would best apply to “local earthquake monitoring efforts” — as in Oklahoma — “where there are high-seismicity rates.” Yeck cautions, though, that earthquake detection is only ever going to be a part of the puzzle. “Estimations of earthquake sizes and accurate event locations are also necessary,” says Yeck. “For the very small events that this technique detects, this will be challenging.”

If this neural network can be used more widely in Oklahoma applied, says Perol, it’ll help seismologists investigate the exact cause of the state’s earthquakes. There’s even some hope that it could predict earthquakes before they occur. This could be done by looking for patterns in the data; for example, finding times when a number of small earthquakes have happened in quick succession, triggering a bigger, potentially damaging quake.

The idea of using AI to predict — not just detect — earthquakes is an exciting one, but it’s not something that the whole seismologist community is confident about. (You can watch the video below for more info.) In Oklahoma at least, prediction isn’t as pressing as detection. But with the help of Perol and his colleagues’ neural network, this important work could get a boost.