Machine Learning Detection of P-Waves in Laboratory Acoustic Emission Events to Understand Deep-Focus Earthquakes

Abstract

The mechanisms of deep-focus earthquakes (DFEQs)—those between 350 and 700 km depth—remain poorly understood due to our inability to directly measure their fault properties in situ. One potential explanation for DFEQ nucleation is the transformational faulting hypothesis, which theorizes mineral transformations initiate the faults. To analyze these structures in more detail, deformation events were conducted in a controlled laboratory environment on 2.1x3.0mm Mg¬2GeO4 samples, the leading mineralogical candidate of the transformational faulting hypothesis. Six orthogonally oriented sensors recorded Acoustic Emission (AE) events to detect P-wave arrival times, returning event-trigger waveform data at 50 MHz and continuous data for 42.4 minutes at 10 MHz. The experiment returned 3,901 event-trigger SAC files and 19,280 continuous SAC files, totaling just over one billion data points.

To analyze this large quantity of raw waveforms, this study introduces machine learning as a tool to automate the detection process. A deep-learning-based detector called EqTransformer (EqT; Mousavi et al., 2020) was trained on global seismic data from the Stanford Earthquake Dataset (Mousavi et al., 2019) and applied to the experimental data to perform P-wave detection and arrival time picking. The short-term goal of this project is to determine the robustness of EqT on microseismic data in both event-trigger and continuous forms. Preliminary results indicate the application of EqT on the event-trigger data was successful. EqT detected 93.4% of the events identified manually, as well as 57.3% additional events missed by the human analysts. The long-term goal is to create a definitive catalog of the AE events that occurred in this experiment, using EqT on the continuous dataset. This could potentially offer key insights into the scaling properties of seismic experiments.