by Harshil Kamdar, PhD student, Harvard University
There is a one hundred percent chance of an earthquake occurring somewhere on Earth today. It may be so light that only the most sensitive instruments on the planet might detect it; it may shake a few windows; or it may be strong enough to cause displacement, injury, or even death. The motivations behind the prediction of earthquakes are pretty clear: save human lives and minimize damage to the surrounding area. On paper, earthquakes seem like something we should at least be able to at least make reasonable guesses about when they might occur — earthquakes are a purely physical system, and thus, unlike economics or elections, are not affected by the “human” element. Moreover, half of the ten deadliest natural disasters in human history (excluding famines and epidemics) have been caused by earthquakes. Given the rapid population growth and urbanization in the last few decades, more than two billion people now live within 250 kilometers from a fault line. Consequently, predicting earthquakes has never been more important. Yet, despite that all the data we have, the physical nature of earthquakes, and the clear motives, we are still unable to predict when they will occur.
Any precise prediction of an earthquake needs to include three different pieces of information: (1) the exact date and time, (2) the physical location, and (3) the magnitude or severity of the earthquake. To date, no major earthquake has been predicted by scientists that included all three of these pieces of information. The key difficulty, according to experts, is actually good news for most of us who are not seismologists: there have been only sixteen recorded earthquakes with a magnitude > 8.5 since 1901. Hence, a lot of the fundamental physics behind major earthquakes is still not fully understood due to the dearth of major earthquakes.
Models so far have not enjoyed a lot of success in predicting major earthquakes due to the enormous complexity of the physics involved. Moreover, GPS sensors today allow us to measure movements of less than 1/10th of a millimeter per year, which is much lower than the typical velocity of a tectonic plate. Consequently, scientists have recently been turning their attention to machine learning models to try to predict earthquakes, given the massive amount of sensor data available but the relative lack of sophistication in the physical models used for prediction.
Geophysicist Paul Johnson and his team are among a handful of groups that are using machine learning to try to demystify earthquake physics and tease out the warning signs of impending quakes. Using sophisticated algorithms similar to those behind recent advances in image and speech recognition, Johnson and his collaborators were able to successfully predict temblors in a laboratory setting. More recently, Johnson and his team claim that they’ve tested their algorithm on slow slip quakes in the Pacific Northwest and show tantalizing results on the potential of using machine learning in earthquake prediction. According to the paper, the algorithm can predict the start of a slow slip earthquake to within only a few days.
Seismologists overwhelmingly agree that even tens of seconds of warning for major earthquakes has the potential to save countless lives. However, most scientists argue that we should be focusing our energy on not just the prediction of earthquakes but also the immediate response in the aftermath of major earthquakes. In terms of infrastructure, for instance, according to Professor Brendan Meade, buildings designed to withstand earthquakes are like bicycle helmets — if there is ever a crash (or an earthquake), they cannot be reused after. The goal of predictive models in seismology, then, is to provide enough time for people to take cover and the community’s first responders to prepare.
While predicting when, where, and how large an earthquake is going to be might be close to an impossible problem to solve, seismologists do make models for earthquake forecasting. Earthquake forecasts can be made, providing the probability that an earthquake of a given size or larger will occur in an area (e.g. California) over a certain timeframe (usually several decades). These forecasts provide information on the likelihood of an event occuring in that timeframe, not certainty as to whether or not it will occur.
We can say with certainty that predictions of weather within a few days are reliable because we understand a lot of the underlying physics. For earthquakes, we emphatically do not know with any certainty how good or reliable the predictions are. The limitation is not the amount of data but that the physical models we have are not very representative of reality. In fact, Professor Meade thinks we may be at the limit of expressing complex physical phenomena with math we think we understand. Time to invent some new math!