Predicting earthquakes: AI to the rescue

When the ground shakes, there are usually consequences - and very often deadly. About 10,000 people die every year during and in the immediate aftermath of earthquakes. But so far, scientists have not had much luck finding ways to accurately predict earthquakes.

"Earthquake prediction is currently impossible," says John Vidale, a seismologist at the University of Washington. "At best, during times of high aftershock danger, low probabilities of danger can be accurately assessed, this effort is termed 'operational earthquake forecasting' in the US."

Prediction is not working, he says, because the factors that determine the time and impact of an earthquake, the distribution and strength of stress deep in the ground, requires measuring a level of detail that is now - and, he adds, "probably forever," unobtainable.

It's not for lack of trying. For decades, scientists have studied foreshocks, changes in groundwater chemistry, odd animal behaviour, electromagnetic disturbances, and so on, all in the hope of find a reliable way to predict an imminent quake. But nothing so far has worked.

At the moment, the best that geologists can do is to determine the particular faults where quakes are likely to take place, and whether they moved in the past. When a fault moves, usually a quake ensues. For instance, a closely monitored fault is the San Andreas Fault in California. Earthquakes there happened in 1857, 1881, 1901, 1922, 1934, and 1966 - so roughly every 22 years. But when geologists assumed a quake would therefore occur between 1988 and 1993, they were way off - it happened only in 2004.

Tremor in the lab

To accurately predict a quake, one has to predict not one but three factors simultaneously: its location, its size, and the time when it'll hit, says Peggy Hellweg, a seismologist at University of California Berkeley. "I can tell you that there will be an earthquake in California tomorrow - there will be many, most of them too small to be felt. The prediction must be for all three characteristics."

She thinks that we cannot predict quakes, because we don't have a broad enough view into the system with enough detail. "If you compare the amount of data we have about quakes with [what we know about] weather, we seismologists are currently collecting data at a level comparable with weather [forecasting] about 100 years ago. We weren't doing a very good job of predicting weather then. Not until we got the much more encompassing view from satellite data."

However, a team at the Los Alamos National Laboratory in New Mexico, led by geophysicist Paul Johnson, seems to have made a step forward - albeit for now only in the lab. They decided to feed a machine-learning algorithm raw data— lots and lots of measurements recorded before, during and after lab-simulated earthquakes. The idea was for the technology to use artificial intelligence and recognise the signs that a lab-created tremor is about to happen, by analysing the sounds emitted by the material under the strain simulating a fault line.

Hellweg, who is not involved in this research, says that Johnson and his group are "currently only working on [predicting] time, and maybe location, but not size [of a quake]." 

A new signal

To create lab earthquakes, the scientists first insert a block between two other blocks, and put in-between a mixture of rocky material, so-called gouge material, to mimic the properties of real faults. Then they begin to pull out the middle block - and just as in a real quake, right before the rupture the gouge material begins to fail, generating specific cracks and sounds in the process. The block then slips quasi-periodically. Geologists believe that this system simulates the behaviours of real quakes.

The team wanted to study whether the sound emitted by the fault could be an indication of when the next slip would occur. To do so, they recorded all the sound from the experiment and fed the data into a machine-learning algorithm - in the hope that the machine would be able to identify specific patterns.

"There is a 'training' phase where the machine-learning algorithm is trained based on knowledge of when a fault may slip and produce an earthquake," explains Johnson. "Then, the machine-learning is applied to data it has never seen before from the same system."

As it turns out, the machine can predict very accurately, just by listening to the acoustic signals emitted by a laboratory fault, when a 'quake' is likely to take place. The researchers think that the algorithm can identify a new signal amidst all the acoustic data — certain creaking and grinding” sounds that scientists used to believe was just noise.

Of course, a lab experiment is very different from real quakes. But in some ways, they are also similar. "We can't be sure we'll ever predict real earthquakes, but we think if progress is to be made, this may be the best possible approach in existence," says Johnson.

The team's next step is to study small ruptures on real faults, such as the San Andreas, where repeating earthquakes occur over relatively short periods. Some seismologists, such as Hellweg, are willing to keep an open mind about the results. She says that "the level of noise in the [Johnson's laboratory] data from other sources is relatively low. So, might there be something there? Yes. How close would we have to be to a source to measure the signals in the real world? Who knows. One challenging question for earthquake prediction systems is - whether it will always work, or at least 90% of the time."

Not everyone is convinced it can work, though. "Paul has great ideas, but I’d be surprised if an unfocused search turns up heretofore unrecognised silver bullets," says Vidale. "Many people have pored over the seconds to days before large earthquakes - and large catalogues of small earthquakes as well - the effort has already been lengthy and intense." But, he admits that he will pay "close attention to the results".