SAN FRANCISCO: Harvard University is working with Google to develop a machine learning-based model to predict aftershock locations after an earthquake. It can help with the deployment of emergency services and help with evacuation plans, said a researcher.
"We worked with experts in machine learning at Google to see if we could apply in-depth learning to explain where aftershocks can occur," wrote Phoebe DeVries, postgraduate student at Harvard, in a Google blog post.
Earthquakes usually occur in series – a first "mainshock" (the event that usually hits the headlines) is often followed by a series of "aftershocks".
Although these aftershocks are usually smaller than the main shock, in some cases they can significantly impede recovery efforts. Although the timing and extent of aftershocks were understood and explained by established empirical laws, predicting the locations of these events has proved to be a greater challenge.
Using deep learning algorithms, the team analyzed a database with information about more than 118 major earthquakes from around the world to predict where aftershocks can occur.
From there, they applied a neural network to analyze the relationships between static stress changes caused by the mainshocks and aftershock locations. The algorithm was able to identify useful patterns.
They developed a system, detailed in the journal Nature, which, although still inaccurate, could predict aftershocks significantly better than arbitrary assignment.
The new model offers opportunities for finding potential physical theories that enable us to better understand natural phenomena, DeVries remarked.
"This machine-acquired insight provides better predictions of aftershock locations and identifies physical quantities that can cause earthquakes during the most active part of the seismic cycle," the researchers explained in the paper.