According to a new study published Wednesday in Nature, neural networks can predict where aftershocks will occur after earthquakes with a higher accuracy than standard techniques.
After a severe earthquake, aftershocks can endanger the survivors and those who want to help them. Now, a group of researchers from Harvard University, the University of Connecticut, and Google have tried to figure this out by training neural networks to predict whether an aftershock will occur at a particular location.
A simple feedforward neuronal network was trained using data from more than 131,000 pairs of major shocks and their aftershocks. The seismic wave patterns are divided into grid cells that measure a fixed volume. The neural network learns to predict whether an aftershock will occur in the center of each grid cell based on the distribution of the stress caused by the earthquake in each region of space.
"By discretizing the volume around each mainshock in this way, aftershock predictions can be formulated as a large-scale binary classification problem, with the goal of measuring every 5 km × 5 km × 5 km grid cell in volume around each mainshock exactly as either" To classify aftershocks "or to include no aftershocks"
The neural network was tested on 30,000 seismic events and was able to determine the possibility of aftershocks more accurately than the method of Coulombic Failure Voltage Change, which is more commonly used
Boffins want machine learning to predict earthquakes
A perfectly accurate classifier has a score of one and a score no better than chance 0.5 The neural network had a classifier v at 0.849 compared to the Coulomb failure voltage change of 0.583.
The Coulomb error voltage change criterion has been widely used to explain aftershocks. We show that the neural network in our test dataset can provide better results in explaining aftershock positions than Coulomb's failure voltage change, "said Phoebe DeVries, a postdoctoral fellow at Harvard University, The Register The model is still on Prototype and does not take into account different types of physical stress so it will not be used in the near future, but development continues and hopes it can be used locally by seismologists and rescuers 19659002] "We see this as one motivating first step. There are many different physical phenomena that can affect the behavior of aftershocks, "said DeVries.
" We focused on static elastic stress changes in this work, but machine learning could be a very powerful tool for linking all kinds of physical phenomena [such as] dynamic stress changes, poroelastic stress changes, existing geological structures, etc. for aftershocks. "®