Predicting quakes in laboratory analogs
To predict how soon an earthquake will strike, geoscientists usually have to rely on the evidence of history—the earthquake’s recurrence time. That’s an average time, though, and the actual time between any two slips on a given fault can vary by years. More accurate forecasting may come from recent advances in computer processing power and in instrumental sensitivity to seismic and acoustic signals. Acoustic precursors to failure appear to be a nearly universal phenomenon in materials and are thought to herald critical stress conditions. But the precursor signals are often small, subtle, and hidden in noise, so the trick is to locate them. Theorist Bertrand Rouet-Leduc of Los Alamos National Laboratory and Cambridge University and his colleagues have developed a technique that extracts the signals from the noise and analyzes their statistical features to infer a material’s instantaneous friction, shear stress, and how likely it is to fail.
To mimic an earthquake, the researchers compressed a grain-coated piston between two steel blocks at a constant normal stress σN, as shown here (top), and drove the piston downward at a constant speed v to induce stick–slip behavior. They recorded sound waves emanating from the shearing layers and fed the data—parsed into short, discrete time windows—into a machine-learning algorithm known as random forest. The algorithm made a sequence of decisions based on each data parcel’s mean, variance, and other statistical features to predict how much time remained before the next slip. The predictions (bottom, blue lines) hew remarkably close to the observed stick–slip events (red lines). Surprisingly, predictions were accurate throughout the entire stick–slip cycle, not only when slip was imminent. The researchers plan to extend their technique to real earthquakes, but they remain cautious about its applicability. Laboratory shear rates (5 µm/s) are orders of magnitude larger than Earth’s (on the scale of centimeters per year). And laboratory temperatures don’t resemble those at Earth’s fault lines. The technique could also be used to predict avalanches, landslides, and the failure of machine parts. (B. Rouet-Leduc et al., Geophys. Res. Lett. 44, 9276, 2017 .)
