Tsunami waves hit the coast in Fukushima Prefecture, Japan, after a powerful earthquake under the North Pacific Ocean.

Global Positioning System (GPS) measurements suggest hours-long precursors to many large earthquakes

A meaningful earthquake prediction must clearly define the expected time, location, and magnitude of a future event (1). Short-term earthquake prediction—that is, the capability to issue a warning from minutes to a few months before a mainshock—is impossible without the existence of an observable and actionable precursor. Thus, a key goal is the discovery of a common preparatory faulting process that can tell us where, when, and how big an impending earthquake is going to be (2). On page 297 of this issue, Bletery and Nocquet (3) present a systematic analysis of changes in horizontal position of approximately 3000 geodetic stations, which were measured by using the Global Positioning System (GPS), near 90 global earthquakes with magnitudes greater than 7. They found that on average, horizontal movements of the stations exponentially accelerated in a direction consistent with slow fault slip near the eventual earthquake nucleation point in the last 2 hours before the earthquake ruptures.

There is a long history of retrospective studies, carried out after a large earthquake has already happened, that suggests a wide variety of possible precursors that could potentially have been used to predict the earthquake (2). In the 1970s, there was some optimism that observations of reliable precursors should be possible with improved geophysical observations, but that optimism waned in subsequent decades. There is a single case of an officially issued prospective prediction of a large earthquake, based largely on the observation of hundreds of small earthquakes, in the days preceding the 1975 Haicheng earthquake in China. Even then, luck apparently played a big role in the identification of these events as foreshocks (4). Nonetheless, many earthquakes are preceded by foreshock sequences (5), and a few of the largest earthquakes of the 21st century, including the 2011 magnitude 9 Tohoku-oki earthquake near the Japan Trench, followed slow faultslip episodes of varying sizes and durations (6). That is, the fault that eventually ruptured and produced earthquake shaking sometimes started moving much more slowly before the mainshock.

In laboratory experiments (7) and in computer models of earthquake ruptures (8), such precursory activity is common. Recently, machine-learning methods were used to study acoustic signals emitted by laboratory faults and successfully predicted the time remaining before the next laboratory quake (9). However, natural foreshocks cannot be distinguished from similar clusters of background seismicity (10), and observed slow preslip events have not appeared different from geodetically measured slow slip transients that occurred without being followed by a large earthquake (11). Overall, earthquake precursors in nature appear to be quite common, but they apparently come in a variety of flavors (6) and have failed to reveal prognostic information that could be used to produce a short-term earthquake prediction (12).

Bletery and Nocquet calculated the sum of observed horizontal displacements of GPS stations in the direction predicted by fault slip at the point where each earthquake rupture nucleated, for each 5-min increment during the 2 days before rupture. Slow fault slip near the eventual nucleation point would contribute to increasing values of this summed function. For each individual earthquake, the signal remains subtle at best, but about half of all earthquakes studied exhibited acceleration in horizontal displacement of nearby geodetic stations in the 2 hours before the mainshocks. The inferred average moment (a measure of the size of a slip event) of these short-duration preslip episodes is equivalent to that of a magnitude 6.3 earthquake.

The authors present several statistical tests to build support for the proposed short-term precursory signal. Only 0.3% of 100,000 applications of their analysis carried out for randomly chosen 48-hour time windows led to the identification of a similarly significant, apparent precursor signal, indicating the specificity of the measurement. Bletery and Nocquet argue that the short duration and exponential acceleration of the preslip signal make the precursory phase they discovered different from quite commonly observed, independent slow slip events.

Although the results of Bletery and Nocquet suggest that there may indeed be an hours-long precursory phase, it is not clear whether such slow-slip accelerations are distinctly associated with large earthquakes or whether they could ever be measured for individual events with the accuracy needed to provide a useful warning. It will be important to fully explore how often similar slow slip episodes occur as false starts, without being followed by earthquakes. There should also be similar analyses of foreshock activity. This will allow evaluation of whether the last-hour slow-slip accelerations can be related to more enduring foreshock activity, which may last weeks to months (5). Where they exist, data from other precise geodetic systems (such as strainmeters, inclinometers, and ocean-bottom pressure sensors) should be reviewed to independently assess the proposed short-term preslip episodes (13). Machine-learning methodologies may be valuable tools to optimally explore these complementary seismic and geodetic data. Most large earthquakes occur in subduction zones, which are largely under the oceans and thus quite distant from GPS monitoring networks. Improving the capability to properly detect offshore slow slip events will require installation of highly accurate and high-rate geodetic measurement systems on the seafloor (14).

The approach of Bletery and Nocquet requires full knowledge of the location and geometry of the mainshocks. Even if it will eventually be possible to detect such preslip events without that information, the short warning window ahead of an imminent earthquake would limit the actions that could be taken to mitigate the impact on people. However, this could possibly be integrated into automated earthquake early-warning systems, which already provide seconds to minutes of warning of shaking to come in some parts of the world, before seismic waves arrive from an earthquake that has already started (15). If it can be confirmed that earthquake nucleation often involves an hours-long precursory phase, and the means can be developed to reliably measure it, a precursor warning could be issued, letting people know that it is time to let go of sharp utensils and get ready to “Drop, Cover, and Hold On,” before the Big One strikes.