Source-linked AI summary
Machine Learning Predicts Laboratory Earthquakes
Bertrand Rouet-Leduc, Claudia Hulbert, Nicholas Lubbers, Kipton Barros, Colin Humphreys, Paul A. Johnson
TL;DR
Forecasting when faults will fail remains an elusive goal in earthquake science. This paper applies machine learning to instantaneous acoustic signals from a laboratory fault, achieving accurate failure-time prediction and identifying a previously unrecognized signal useful throughout the quake cycle.
Problem
The paper addresses the longstanding challenge of forecasting fault failure times, including whether laboratory labquake failure can be predicted.
Method
Machine learning analyzes instantaneous acoustic-signal characteristics, including signal-amplitude distribution statistics, without using signal history.
Results
The random-forest model predicts laboratory fault failure time with an R2 value of 0.89 and identifies a previously unrecognized predictive signal throughout the slip cycle.
Takeaways & Limitations
The results provide information on laboratory fault-failure timing and may support new insights into fault physics and bounds on failure times.
Takeaways & Limitations
The reported scope concerns earthquake timing rather than earthquake magnitude.
Abstract
from arXiv · showhide
Forecasting fault failure is a fundamental but elusive goal in earthquake science. Here we show that by listening to the acoustic signal emitted by a laboratory fault, machine learning can predict the time remaining before it fails with great accuracy. These predictions are based solely on the instantaneous physical characteristics of the acoustical signal, and do not make use of its history. Surprisingly, machine learning identifies a signal emitted from the fault zone previously thought to be low-amplitude noise that enables failure forecasting throughout the laboratory quake cycle. We hypothesize that applying this approach to continuous seismic data may lead to significant advances in identifying currently unknown signals, in providing new insights into fault physics, and in placing bounds on fault failure times.
Affiliations:
The authors are affiliated with Los Alamos National Laboratory, the University of Cambridge, and Boston University.
- Affiliations:: The affiliations span Los Alamos National Laboratory, the University of Cambridge, and Boston University.Los Alamos affiliations include the Theoretical Division and CNLS and the Geophysics Group; Cambridge affiliation is in Materials Science and Metallurgy, and Boston University affiliation is in Physics.
Introductory Paragraph:
The approach may provide new insights into fault physics and place bounds on fault failure times.
- The approach may provide new insights into fault physics and place bounds on fault failure times.
Main Text:
Machine learning predicts laboratory fault-failure times from instantaneous acoustic-signal characteristics without using signal history. The approach achieves high accuracy throughout the quake cycle, identifies a previously overlooked signal, and generalizes across experimental conditions.
- Prediction method: Instantaneous acoustic-signal characteristics predict time remaining before failure without using the system’s history.The prediction uses information from a single time window and listens to the signal currently emitted by the system.
- Prediction performance: Random-forest predictions achieve an R2 of 0.89, compared with 0.3 for periodicity-based predictions.The model predicts failure accurately throughout the entire laboratory earthquake cycle, not only when failure is imminent.
- Predictive signals: Signal-amplitude distribution statistics forecast failure, with variance strongest early and fourth-moment and threshold statistics becoming predictive near failure.These later outlier statistics respond to impulsive precursor acoustic emissions observed as the material approaches failure.
- Novel signal: A minute, increasing acoustic-emission signal long before failure was previously assumed to be noise and resembles tremor associated with slow slip.The findings indicate that the system emits increasing energy throughout the stress cycle before abruptly releasing accumulated energy during a slip event.
- Generalization: Model predictions retain accuracy across load levels and generalize to aperiodic fault cycles, including conditions the random forest has never seen.This suggests the time-series signal captures quantitative frictional-state information relevant to the timing of the next slip event.
- Implications: The study demonstrates accurate failure forecasts from instantaneous acoustic analysis and motivates applying machine learning to continuous seismic signals.The approach may reveal unidentified signals associated with undiscovered fault physics, although laboratory experiments cannot capture all complex Earth-rupture physics.
Figures
The figures present Cascadia’s earthquake context and illustrate a Random Forest method that uses acoustic-emission data to predict time remaining before laboratory fault failure. They identify laboratory tremor and impulsive acoustic emissions as signals associated with different stages of the failure cycle.
- Earthquake context: Cascadia subduction and earthquakes larger than M8.0 are illustrated alongside estimated event timing and time remaining before the next event.The figure depicts the Juan de Fuca Plate subducting beneath the North American plate near Seattle and uses turbidites to estimate past earthquakes.
- Prediction method: The Random Forest predicts time remaining before failure from acoustic-emission dynamic-strain data extracted through moving time windows.Shear-stress drops define laboratory failures and the target time remaining, while each window produces feature values such as variance and kurtosis.
- Prediction results: 90 percent of the Random Forest trees provide forecasts within the plotted 5th- and 95th-percentile bounds.The figure compares actual time before failure with predictions obtained from successive time windows and highlights aperiodic slip behavior in an inset.
- Physics of failure: The Random Forest identifies laboratory tremor when failure is distant and classic impulsive acoustic emission immediately before failure.The figure presents these as two signal classes used to predict failure, with laboratory tremor offering precise prediction of the next failure time.