Aftershocks occur as a response to stress changes produced by a large earthquake. Unlike their maximum magnitudes and frequencies, which are well described by empirical laws, predicting and explaining where aftershocks will occur is rather difficult.
Previously, a stress-transfer criterion – known as Coulomb failure stress change – was commonly used to explain the spatial distribution of aftershocks, but it didn’t work in all cases.
In a paper published in the journal Nature, researchers have now used a machine-learning approach to establish a stress-based law that can more accurately predict and explain the locations of more than 30,000 aftershocks compared to the conventional Coulomb metric.{%recommended 5750%}
The US-based team, led by Phoebe DeVries from the University of Connecticut, used more than 131,000 additional mainshock-aftershock pairs to first train the artificial neural network.
“Aftershock forecasting is very well-suited for machine learning in many ways,” explains DeVries.
“There are a lot of data – thousands of aftershocks take place in the wake of large earthquakes – and there are a lot of different physical phenomena that may affect aftershock behaviour.”
This is not the first time artificial neural networks have been trained to predict the spatial distributions of aftershocks. However, it is the first to expand the research beyond a single earthquake event.
The new findings also offer additional insight into potential earthquake-triggering mechanisms. The locations identified by the deep-learning approach could be more aptly explained by three other well-known stress criterions — including one that had not previously been proposed as a potential factor.
DeVries says that these findings are encouraging but, “there is a long way to go before these kinds of forecasts could be useful in any operational sense”.
“The core idea here is that machine learning approaches may offer significant gains in terms of forecasting complex earth and environmental systems,” she adds.
“With the vast amounts of satellite and earth-based data available today, machine learning techniques will continue to be applied to better predict the evolution of our natural world.”