Predicting earthquakes is the holy grail of seismology. After all, quakes are deadly precisely because they’re erratic—striking without warning, triggering fires and tsunamis, and sometimes killing hundreds of thousands of people. If scientists could warn the public weeks or months in advance that a large temblor is coming, evacuation and other preparations could save countless lives.

So far, no one has found a reliable way to forecast earthquakes, even though many scientists have tried. Some experts consider it a hopeless endeavor. “You’re viewed as a nutcase if you say you think you’re going to make progress on predicting earthquakes,” says Paul Johnson, a geophysicist at Los Alamos National Laboratory. But he is trying anyway, using a powerful tool he thinks could potentially solve this impossible puzzle: artificial intelligence.

Researchers around the world have spent decades studying various phenomena they thought might reliably predict earthquakes: foreshocks, electromagnetic disturbances, changes in groundwater chemistry—even unusual animal behavior. But none of these has consistently worked. Mathematicians and physicists even tried applying machine learning to quake prediction in the 1980s and ’90s, to no avail. “The whole topic is kind of in limbo,” says Chris Scholz, a seismologist at Columbia University’s Lamont–Doherty Earth Observatory.

But advances in technology—improved machine-learning algorithms and supercomputers as well as the ability to store and work with vastly greater amounts of data—may now give Johnson’s team a new edge in using artificial intelligence. “If we had tried this 10 years ago, we would not have been able to do it,” says Johnson, who is collaborating with researchers from several institutions. Along with more sophisticated computing, he and his team are trying something in the lab no one else has done before: They are feeding machinesraw data—massive sets of measurements taken continuously before, during and after lab-simulated earthquake events. They then allow the algorithm to sift through the data to look for patterns that reliably signal when an artificial quake will happen. In addition to lab simulations, the team has also begun doing the same type of machine-learning analysis using raw seismic data from real temblors.

This is different from how scientists have attempted quake prediction in the past—they typically used processed seismic data, called “earthquake catalogues,” to look for predictive clues. These data sets contain only earthquake magnitudes, locations and times, and leave out the rest of the information. By using raw data instead, Johnson’s machine algorithm may be able to pick up on important predictive markers.

Johnson and collaborator Chris Marone, a geophysicist at The Pennsylvania State University, have already run lab experiments using the school’s earthquake simulator. The simulator produces quakes randomly and generates data for an open-source machine-learning algorithm—and the system has achieved some surprising results. The researchers found the computer algorithm picked up on a reliable signal in acoustical data—“creaking and grinding” noises that continuously occur as the lab-simulated tectonic plates move over time. The algorithm revealed these noises change in a very specific way as the artificial tectonic system gets closer to a simulated earthquake—which means Johnson can look at this acoustical signal at any point in time, and put tight bounds on when a quake might strike.

For example, if an artificial quake was going to hit in 20 seconds, the researchers could analyze the signal to accurately predict the event to within a second. “Not only could the algorithm tell us when an event might take place within very fine time bounds—it actually told us about physics of the system that we were not paying attention to,” Johnson explains. “In retrospect it was obvious, but we had managed to overlook it for years because we were focused on the processed data.” In their lab experiments the team looked at the acoustic signals and predicted quake events retroactively. But Johnson says the forecasting should work in real time as well.

Of course natural temblors are far more complex than lab-generated ones, so what works in the lab may not hold true in the real world. For instance, seismologists have not yet observed in natural seismic systems the creaking and grinding noises the algorithm detected throughout the lab simulations (although Johnson thinks the sounds may exist, and his team is looking into this). Unsurprisingly, many seismologists are skeptical that machine learning will provide a breakthrough—perhaps in part because they have been burned by so many failed past attempts. “It’s exciting research, and I think we’ll learn a lot of physics from [Johnson’s] work, but there are a lot of problems in implementing this with real earthquakes,” Scholz says.

Johnson is also cautious—so much so that he hesitates to call what he is doing “earthquake prediction.” “We recognize that you have to be careful about credibility if you claim something that no one believes you can do,” he says. Johnson also notes he is currently only pursuing a method for estimating the timing of temblors, not the magnitude—he says predicting the size of a quake is an even tougher problem.

But Scholz and other experts not affiliated with this research still think Johnson should continue exploring this approach. “There’s a possibility it could be really great,” explains David Lockner, a research geophysicist at the U.S. Geological Survey. “The power of machine learning is that you can throw everything in the pot, and the useful parameters naturally fall out of it.” So even if the noise signals from Johnson’s lab experiments do not pan out, he and other scientists may still be able to apply machine learning to natural earthquake data and shake out other signals that do work.

Johnson has already started to apply his technique to real-world data—the machine-learning algorithm will be analyzing earthquake measurements gathered by scientists in France, at Lawrence Berkeley National Laboratory and from other sources. If this method succeeds, he thinks it is possible experts could predict quakes months or even years ahead of time. “This is just the beginning,” he says. “I predict, within the next five to 10 years machine learning will transform the way we do science.”