 |
Center for Nonlinear Studies |
Forecasting failure is the elusive, holy grail in diverse domains that include earthquake physics, materials science and many engineering applications. Due to the highly complex physics of material failure, the goal appears out of reach; however, recent advances in instrumentation sensitivity, instrument density and data analysis show promise toward forecasting failure times. In particular, observations of seismic precursors to earthquakes that occur naturally or are modulated by Earth tides suggest that at least in some instances, the time interval of an upcoming failure can be bounded, if imprecisely. Here we show that we can predict frictional failure times (‘labquakes’) with great accuracy in laboratory shear experiments. This advance is made possible by applying Random Forest (RF) machine learning to the continuous time series recorded by a single accelerometer listening to the experiment. The RF is trained applying a number of statistical data features over a time interval over which a number of labquakes occur. Remarkably, during testing we find that the RF predicts upcoming failure time immediately following a labquake, based only on a short time window of data—a ‘now ’ prediction. The predicted time improves as failure is approached, as other data features add to prediction. Strikingly, the features correspond to signals identified as noise preceding the analysis, but that are in fact creep-like signals resembling Earth tremor that occurs deep in faults. Impulsive acoustic emission precursors typically observed take place in tandem with ongoing tremor immediately preceding failure. We anticipate that, if similar signals exist in Earth faults and in other materials approaching failure, great advances can be made in failure prediction. Moreover, as we see in the laboratory experiment, the RF approach may identify new acoustic emission and seismic signals that are associated with unknown and unexpected physics.
Host: Chris Neale