Predicting fault stress level and stress drop using seismo-mechanical and statistical features derived from acoustic signals in laboratory stick-slip friction experiments and assesing feature importance via the derived models.

Earthquake prediction relies on identification of distinctive patterns of precursory parameters that might precede a large earthquake. However, these patterns are typically not reliably observed in the field. Laboratory stick-slip experiments provide an analog of seismic cycle observed in nature in fully controlled conditions with the associated Acoustic Emission (AE) activity reproducing basic characteristics of seismicity preceding and following the large lab earthquake. Recent laboratory studies showed that the deployment of Machine Learning/Artificial Intelligence techniques has lead to new state of the art results in lab earthquake prediction on smooth faults while using simple statistical features derived from raw AE signals and AE-derived catalogs. However, not enough work has been done on explainability of earthquakes preparatory process on rough faults by leveraging deep learning techniques. In this work we attempt to mitigate this gap and analyze/grade a pool of  explainable seismo-mechanical features through the eyes of neural networks.

We used AE data from three laboratory stick-slip experiments performed in triaxial pressure vessel on Westerly Granite samples. Samples were first fractured at 75MPa confining pressure creating rough fault surfaces. The following stick-slip experiments were performed at constant displacement rate. The experimental procedure led to an extremely complex slip pattern composed of large and small slips of the whole surface, as well as the confined slips highlighted only with AE data and no externally measured slip. The AE catalog was used to extract temporal evolution of 16 seismo-mechanical and statistical features characterizing evolution of stress and damage in response to the axial stress change. The feature pool included clearly physically interpretable parameters such as AE rates, b-value, fractal dimension, AE localization, clustering and triggering properties, and features characterizing the variability of local stress field. 

We apply explainable AI techniques to identify what features are more important to forecast  axial stress and stress drop.  Our feature ranking and importance evaluation with the help of neural networks can serve as an indicator as to what research directions are more promising to take for further feature engineering efforts with an emphasis on explainability of earthquake phenomena.