Creative Commons

Earthquake prediction, the coveted holy grail of earthquake science continues baffle Earth scientists (Johnson et al., 2021). Could we make progress by using the huge expertise and innovation of the ML’s community? We used a popular Machine Learning competition platform, Kaggle, to join the global ML community in a competition, to gain access in their dataset so as to build and enhance data analysis algorithms for a forecasting challenge including experimental laboratory earthquake data. Participants were under the challenge of predicting the remaining time until the next earthquake in consecutive laboratory earthquakes with a relatively small portion of the laboratory seismic data. We use machine learning to find hidden signals that anticipate earthquakes in datasets from shear laboratory studies. We demonstrate that machine learning can accurately forecast the time before a defect fails by listening to the acoustic signals of a laboratory fault. These forecasts do not take into account the acoustic signal's history and are entirely based on its instantaneous physical characteristics. Unexpectedly, machine learning discovers a fault zone signal that was previously assumed to be of low utility noise and authorizes failure prediction all across the laboratory earthquake cycle. (Rouet-Leduc et al., 2017). We conclude that this signal comes from continuous movements of the fault furrow as the fault blocks shift. We believe that using this method with continuous seismic data may result in important advancements in recognizing previously unrevealed signals, new understandings of the physics of faults, and the establishment of limits on fault downtime.