“Deep learning of aftershock patterns following large earthquakes,” by Phoebe DeVries, Fernanda Viegas, Martin Wattenberg, and Brendan Meade, appeared in Nature in August 2018. It asked whether a neural network could predict the spatial pattern of aftershocks, the smaller quakes that follow a large one, better than the standard physics-based rule.
The team trained a network on more than 131,000 mainshock-aftershock pairs and tested it on more than 30,000 held-out pairs. The network forecast aftershock locations more accurately, by a common area-under-curve measure, than the classical Coulomb failure-stress approach, and the authors highlighted which stress quantities the model relied on as a possible physical clue.
The paper became a notable case study in the limits and pitfalls of applying deep learning to science. A 2019 follow-up argued that a far simpler model, in effect a single neuron, matched the deep network’s performance, suggesting the headline benefit came less from deep learning than from the choice of inputs. The original authors responded, and the exchange is now cited as a cautionary tale about strong claims and proper baselines.
For a general reader, this entry is valuable precisely because of the controversy: it shows AI for science working in a new domain while also illustrating how easy it is to overstate what a complex model adds, and why careful comparison against simple baselines matters.