Phoebe DeVries is currently a post-doctoral fellow at Harvard University. Her research focuses on computational modeling and machine learning for earthquake science applications. She holds a Ph.D. from Harvard University in Earth and Planetary Sciences, a Masters degree from Cambridge University in glaciology, and a B.A. from Harvard College in Applied Mathematics. She will be starting at the University of Connecticut as an assistant professor in January 2019.
Aftershock sequences represent ubiquitous observations of earthquake triggering. Maximum aftershock sizes and time decay can be well described by empirical laws (e.g., Bath’s law and Omori’s law), but explaining and forecasting the spatial distribution of aftershocks has proven more difficult. Static elastic Coulomb failure stress change is perhaps the most commonly invoked framework to explain the spatial distributions of aftershocks, but its applicability has been disputed in some cases. Here we use neural networks to learn aftershock location, density, and magnitude patterns. A fully connected neural network trained on 131,000+ mainshock-aftershock pairs can explain aftershock locations in an independent testing data set of 30,000+ mainshock-aftershock pairs more accurately (AUC = 0.849) than static elastic Coulomb failure stress change (AUC = 0.583). In contrast to the common assertion that deep learning produces ‘black box’ results, the learned aftershock patterns are also physically interpretable: maximum shear stress change and the von-Mises yield criterion, a scaled version of the second invariant of the deviatoric stress tensor, each explain >98% of the variance in the neural network aftershock location forecast. These and similar machine learning driven approaches have the potential to usher in a new generation of predictive modeling in the earth sciences.