A machine learning approach has been used to identify a stress-based law that can forecast the pattern of aftershock locations following large earthquakes. The findings are reported in this week’s Nature.
Aftershocks are a response to seismic stress changes generated by large earthquakes, and empirical laws exist that describe their size and frequency. However, explaining and forecasting the location of aftershocks has proved more difficult. Previously, a factor called Coulomb failure stress change (based on the transfer of stress during an earthquake to surrounding material) has been used to explain aftershock locations, but its applicability has been disputed.
Phoebe DeVries and colleagues trained a neural network using data from more than 131,000 pairs of earthquakes and aftershocks. They found that their network was able to identify and explain the pattern of aftershock locations in an independent dataset of more than 30,000 earthquake-aftershock pairs more accurately than can Coulomb failure stress change. The authors argue that the findings highlight how deep learning approaches can lead to improved aftershock forecasts and provide insights into the mechanisms of earthquake triggering.