Dataset for: Readapting PhaseNet to Laboratory Earthquakes: AEsNet, a Robust Acoustic Emission Picker Illuminating Seismic Signatures of Different Fault Gouge Materials

This dataset supports the study "Readapting PhaseNet to Laboratory Earthquakes: AEsNet, a Robust Acoustic Emission Picker Illuminating Seismic Signatures of Different Fault Gouge Materials" and includes approximately 30,000 manually picked acoustic emission (AE) waveforms. AEs are tiny elastic waves generated by micro-ruptures occurring when a material is deformed. In laboratory shear experiments, where rock samples are subjected to shear forces, these micro-ruptures mimic small-scale earthquakes, offering crucial insights into the micromechanics of different materials.

The waveforms were recorded during laboratory shear experiments performed using the biaxial apparatus, BRAVA2, located in the Rock Mechanics and Earthquake Physics Laboratory at Sapienza University of Rome. These experiments were specifically designed to study frictional instabilities in fault gouge materials. Two different materials—Min-U-Sil quartz gouge and glass beads—were tested under varying experimental boundary conditions to capture a wide range of behaviors.

This dataset serves as a benchmark for training and evaluating AEsNet, a deep learning-based AE detection model adapted from the PhaseNet model (Zhu & Beroza, 2019), originally developed for natural earthquake phase detection. The study demonstrates AEsNet's superior performance compared to traditional AE picking methods, explores the transferability of machine learning models across different fault gouge materials and experimental conditions, and investigates the distinct frequency characteristics of AEs to better understand their connection to the microphysical processes in granular materials.

The dataset includes high-precision, labeled AE events, enabling the development of robust machine learning models for automated AE detection. Comprehensive metadata accompanying the dataset describe experimental setups, material properties, and boundary conditions. The data are formatted in HDF5, ensuring compatibility with machine learning workflows. In addition to the data, the repository includes the pre-trained AEsNet models and a Jupyter notebook with scripts and functions for reproducing the figures presented in the study. The full codebase for retraining new models and applying pre-existing ones will be avalibale at GitHub repository upon publication of the associated paper.

This dataset and its associated tools provide a robust foundation for advancing research into fault mechanics in controlled experimental settings and offer scalable solutions for analyzing acoustic emissions in both laboratory and natural environments.