A machine-learning study that analyzed hundreds of thousands of earthquakes defeated the standard method in predicting the location of aftershocks.

Scientists say the work offers a new way to investigate how changes in soil stress, such as those occurring during a major earthquake, cause the earthquakes that follow. It can also help researchers develop new methods for assessing seismic risk.

"We have really just scratched the surface of what mechanical learning can do for aftershock prediction," said Phoebe DeVries, a seismologist at Harvard University in Cambridge, Massachusetts. She and her colleagues report their findings on August 29th Nature.

Aftershocks occur after the big earthquake and they can be just as damaging – or much more – than the first shock. A magnitude-7.1 earthquake near Christchurch, New Zealand, in September 2010, nobody killed, but a magnitude-6.3 aftershock, which followed more than 5 months later and came closer to the city center, resulted in 185 deaths.

Seismologists can generally predict how great aftershocks will be, but they are struggling to predict where the earthquakes will occur. Until now, most scientists have used a technique that calculates how an earthquake changes the stress in nearby rocks and then predicts how likely it is that this change would result in an aftershock at a particular location. This stress-failure method can successfully explain aftershock patterns for many major earthquakes, but it does not always work.

Large amounts of data are available on previous earthquakes and DeVries and her colleagues have decided to use them to come up with a better prediction method. "Machine learning is such a powerful tool in such a scenario," says DeVries.

Neural networks

The scientists looked at more than 131,000 earthquakes in mainshocks and aftershocks, including some of the most powerful vibrations in recent history, such as the devastating scale-9.1 event that struck Japan in March 2011. The researchers used this data to create a neural network. training that was modeled was a grid of cells, 5 kilometers to one side, around each main shock. They told the network that an earthquake had occurred and gave data on how the stress in the middle of each grid cell changed. Then the scientists asked for the chance that each grid cell would generate one or more aftershocks. The network treated each cell as its own small, isolated problem to solve, instead of calculating how stress successively rippled through the rocks.

When the researchers tested their system for 30,000 main shock aftershock events, the prediction of the neural network more accurately predicted the locations of the aftershock than the usual stress-failure method. Perhaps more importantly, DeVries says, the neural network also gave an indication of some of the physical changes that possibly took place in the ground after the main shock. It pointed to certain parameters as potentially important, which describe stress changes in materials such as metals, but which researchers do not often use to study earthquakes.

The findings are a good step to investigate aftershocks with fresh eyes, says Daniel Trugman, a seismologist at the Los Alamos National Laboratory in New Mexico. "The algorithm for machine learning teaches us something fundamental about the complex processes that underlie the earthquakes," he says.

The latest study will not be the last word about aftershock predictions, says Gregory Beroza, a geophysicist at Stanford University in California. For example, it does not take into account a kind of stress change that occurs when seismic waves travel through the earth. But "this document should be seen as a new perspective on the activation of aftershocks," he says. "That's important, and it's motivating."

This article was accepted with permission and was first published on August 29, 2018.