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
Earthquake recurrence can indicate approaching fault failure, but offers estimates with limited precision. Here, random forests use instantaneous acoustic emissions from a laboratory fault to predict failure time, attaining R² = 0.89 versus 0.3 for periodicity alone.
Problem
Earthquake recurrence estimates can indicate approaching fault failure, but have limited precision, leaving accurate failure-time forecasting an elusive goal.
Source: Main Text · pp. 2–3Method
The study applies random-forest machine learning to moving windows of fault acoustic emissions to predict time remaining before failure from instantaneous signal features, not signal history.
Source: Main Text · pp. 5–6Results
R² was 0.89 for random-forest time-to-failure predictions, versus 0.3 for a periodicity-only naïve model, and predictions remained accurate throughout the labquake cycle.
Source: Main Text · p. 6Takeaways & Limitations
Acoustic-signal variance and higher-order moments carry quantitative information about the instantaneous frictional and shear-stress state.
Source: Main Text · p. 8Whether this approach can provide earthquake timing information from continuous seismic or other geophysical data remains unresolved.
Source: Main Text · p. 9
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.
Introductory Paragraph: · Main Text:
In laboratory faults, a random-forest model forecast time to failure from instantaneous acoustic windows, achieving R2=0.89 versus 0.3 for a periodicity-only model throughout the labquake cycle. The analysis also revealed low-amplitude acoustic activity previously treated as noise, suggesting a potential route to studying failure signals in Earth.
Introductory Paragraph:
- Recurrence intervals and seismic catalogs provide limited timing precision, while improved instrumentation has revealed slow slip, tremor, and potentially small or unidentified acoustic/seismic precursors.The paper notes that catalog records span only a century and that recurrence is not constant; the forecasting commission also warned of methodological limitations and possible bias toward positive precursor observations.Source: Main Text · pp. 2–3
Main Text:
- The study applies machine learning to a two-fault laboratory system with gouge, where quasi-periodic labquakes are preceded by impulsive acoustic emissions as stress builds toward failure.The authors note that laboratory and Earth conditions differ, although the laboratory Gutenberg-Richter relation falls within observed Earth bounds and suggests some fault-friction physics scale.Source: Main Text · pp. 3–4
- A random forest used statistical features from local acoustic-emission windows to predict time remaining before failure, with each prediction based only on the current window and not signal history.The analysis computed approximately 100 candidate features, including mean, variance, kurtosis, and autocorrelation.Source: Main Text · pp. 5–6
- R2 was 0.89 for random-forest time-to-failure predictions, compared with 0.3 for a naïve model based exclusively on event periodicity; predictions were accurate throughout the labquake cycle.The result indicates that the fault progresses toward failure continuously, not only when impulsive precursors are observed near failure.Source: Main Text · p. 6
- The model identified persistent, low-amplitude acoustic modulations previously assumed to be noise, which increased in amplitude and energy as failure approached.The authors posit that continuous grain motions during slow block slip produce this chattering signal, before accumulated energy is released in a slip event.Source: Main Text · p. 7
- Predictions retained accuracy across applied load levels and when a shortened recurrence time disrupted cycle periodicity, supporting generalization to unseen conditions and aperiodic fault cycles.The authors interpret signal variance and higher-order moments as fingerprints of instantaneous friction and shear-stress state.Source: Main Text · pp. 7–8
- The authors argue that instantaneous acoustic analysis reveals signals missed by catalog-based approaches and could motivate machine-learning analysis of continuous Earth seismic data, while acknowledging that laboratory-to-Earth scaling remains a substantial leap.They propose testing whether faults with small repeating earthquakes, including near Parkfield, emit analogous chattering signals detectable by borehole or surface instruments.Source: Main Text · pp. 8–9
Figures
The figures show that Random Forest predictions from acoustic-emission windows forecast laboratory failure times, including aperiodic slip, and identify signals associated with both distant and imminent failure.
Figure 1
- The Cascadia illustration depicts Juan de Fuca Plate subduction beneath North America and estimates of past >M8.0 earthquakes from turbidites; it describes the region as locked and due for a megaquake and tsunami.The estimates cover the last 10,000 years.Source: Figure 1 · p. 14
Figure 2
- A Random Forest averages 1,000 decision-tree predictions from acoustic-signal features in moving windows to forecast time remaining before failure.The depicted test data are entirely independent of training data, and predictions agree with actual remaining times.Source: Figure 2 · p. 15
Figure 3
- Trained on ~150 seconds of data and tested on the following ~150 seconds, the Random Forest forecasts aperiodic slip with each prediction independent of past and future history.The blue forecast and red actual time-to-failure curves are shown with 90% intervals.Source: Figure 3 · p. 16
Figure 4
- The Random Forest identifies laboratory tremor when failure is distant and impulsive acoustic emissions just before failure.Laboratory tremor offers predictive capability for the next failure time; the impulsive signals are also observed before brittle failure in materials and some earthquakes.Source: Figure 4 · p. 17