 |
Center for Nonlinear Studies |
Forecasting failure is an elusive Holy Grail in diverse domains that include earthquake physics, materials science, and numerous engineering applications. Due to the highly complex physics of material failure, the goal may appear out of reach; however, recent advances in instrumentation sensitivity, instrument density, and data analysis show promise toward forecasting failure times in these diverse applications. Here we show in laboratory shear experiments that we can predict frictional failure times (‘labquakes’) with great accuracy, for both 'regular' and slow slip events. This advance is made possible by analyzing the continuous time series—the acoustic emission —recorded from the shearing process using machine learning. Remarkably, we find that our approach can predict the upcoming failure time based only on a short time window of data—a ‘now’ prediction—anywhere in the labquake cycle. Our algorithm also tells us which variables are relevant to forecasting failure. Strikingly, the features which give long-term forecasting information correspond to signals previously thought to be low-amplitude noise, but that are in fact signals resembling Earth tremor that occurs deep in faults. As failure becomes imminent, impulsive acoustic emission precursors take place in tandem with ongoing tremor. We anticipate that significant advances can be made in failure forecasting if similar signals exist in Earth faults as well as in other materials approaching failure. Moreover, our result demonstrates that machine learning approaches can identify new acoustic and seismic signals that are associated with unknown and unexpected physics.
Host: Chris Neale