Probing slow and fast slip events in the laboratory applying machine learning

Earthquake hazard assessment and forecasting is an important problem in earthquake physics. Laboratory friction experiments coupled with ultrasonic measurements are a common approach to illuminating the physical processes that occur during the nucleation stage of laboratory slip events. In previous works, we show that machine learning (ML) can infer physical properties of a laboratory fault zone based purely on statistical features of the acoustic signal. Specifically, we show that a ML model is able to infer the instantaneous shear stress, time remaining until the next slip event and the size of the next slip event. Here, we build on this work by showing that ML can inform us of the instantaneous shear stress for both slow and fast slip events over a range of shearing velocities and normal stresses. We have conducted a suite of friction experiments in a bi-axial loading frame using a double-direct shearing configuration. We shear 3 mm thick layers of quartz powder (min-u-sil) with a nominal contact area of 10 x 10 cm². Experiments were carried out over a range of normal stresses between 6-10 MPa and shearing velocities between 3-17 μm/s. In order to produce both slow and fast slip events, we control the loading stiffness and match it to the critical frictional weakening rate with slip k_c. We record acoustic emission time series data continuously throughout the experiment at 4 MHz from two broad-band piezoceramic sensors. The sensors are embedded inside a steel loading block and positioned adjacent to the fault zone. Our machine learning approach consists of computing 90 statistical features of the acoustic time series signal using a moving window approach. Preliminary results show that a ML model can be constructed at each velocity to infer the instantaneous shear stress on the fault. In addition, we show that the same predictions can be made by implementing the ML model with only the acoustic variance. This relationship between shear stress and variance suggests that it is possible to infer what stage the laboratory fault is in based purely on the acoustic variance. Ongoing work involves finding a scaling relationship between the variance and shear stress, such that, a single ML model can be constructed to infer the shear stress at any velocity and normal stress.