Google and Harvard use machine learning to predict aftershocks
Researchers from Harvard and Google have published a paper in Nature showing how they used deep learning to predict earthquake aftershock locations, The Verge reported.
Their method is significantly more accurate than existing models, the researchers found.
The scientists used 131,000 data points which consist of an earthquake and corresponding aftershock – a mainshock-aftershock event – to train a neural network.
It was then tested on 30,000 data points and its accuracy measured against the model known as “Coulomb failure stress change.”
Coulomb failure stress change was able to predict aftershocks with a 58.3% accuracy, while the deep neural network the researchers developed achieved 84.9% accuracy.
It was able to do this by taking the “Von Mises Yield Criterion” into account. The scientists said it is used in metallurgy to predict when materials will start breaking under stress, but hasn’t been popular in earthquake science.
The machine learning model currently only considers static stress – permanent changes to the ground. However, aftershocks can also be caused by dynamic stress, such as rumblings in the ground that happen later.
This machine learning method is also currently too slow to work in real time, said the researchers.
Now read: Google officially launches Titan Security Key
Loading ...