Typically, major earthquakes are followed up by aftershocks, which are usually smaller, less intense tremors, but also occasionally quite powerful in their own right. These quakes could continue not just for a day or two, but for multiple weeks or even months. But while they may be easy for seismologists and other scientists to predict in a fairly accurate fashion, researchers from one of America’s most prestigious institutions and one of the world’s top tech companies have teamed up to create an artificial intelligence system that could predict aftershocks with much greater accuracy than ever before.
Earlier this week, researchers from Harvard University and Google published a study in the journal Nature that details how deep learning techniques could allow for extremely accurate earthquake aftershock predictions when compared to existing systems. The methodologies involved training a neural network to deduce whether patterns existed in a database of over 131,000 so-called “mainshock-aftershock” combos where a “main” earthquake would be followed by an aftershock. The network was then tested out on a smaller, separate database containing 30,000 similar events.
According to Venture Beat, the former database included information from several historically intense and destructive tremors, including the 1989 and 1994 California earthquakes in the San Francisco Bay Area and Northridge respectively, the 2004 Sumatra earthquake, and the 2011 Japan earthquake.
Prior to the new study, the Coulomb failure stress change model was considered to be the most reliable and accurate model for aftershock prediction. According to the Verge, the Coulomb model scored a 0.583 in a test determining the accuracy of predictions, with 0 being the least accurate, 0.5 meaning 50-50 accuracy, and a score of 1 representing perfect accuracy. The Harvard and Google model proved to be far more accurate, as it registered a score of 0.849.
“We found that after feeding these model stress changes into the neural network, the neural network could sort of predict aftershock locations in the testing dataset more accurately than the sort of baseline Coulomb failure stress change criterion that’s used a lot in studies of aftershock locations,” Harvard University Department of Earth and Planetary Sciences researcher and study co-author Phoebe DeVries told Venture Beat.
The Verge quoted another study co-author, Harvard professor of Earth and planetary sciences Brendan Meade, who said that the results could represent a major breakthrough in earthquake aftershock prediction.
“There are three things you want to know about earthquakes. When they are going to occur, how big they’re going to be and where they’re going to be. Prior to this work, we had empirical laws for when they would occur and how big they were going to be, and now we’re working the third leg, where they might occur.”
While it can be difficult to predict main earthquakes and/or aftershocks with accuracy due to all the factors that could skew analysis and eliminate the possibility of a one-size-fits-all solution for all earthquakes, the Verge wrote that the use of AI in the Harvard/Google study was noteworthy because the system was able to detect patterns that scientists had previously failed to recognize.
Despite its initial promise, the study had its share of limitations as it only covered aftershocks caused by static stress, or permanent changes to the ground in an area affected by an earthquake, and because dynamic stress, or tremors that occur after the fact, could also trigger aftershocks. Furthermore, the Verge noted that the AI system is still “too slow to work in real-time,” which could reduce its usefulness because most aftershocks happen within the first day after the main quake.