Machine learning bridges microslips and slip avalanches of sheared granular gouge

Understanding the origin of stress avalanche of fault gouges may offer deeper insights into many geophysical processes such as earthquakes. Microslips of sheared granular gouges were found to be precursors of large slip events, but the documented relation between local and global avalanches remains largely qualitative. We examine the stick-slip behavior of a slowly sheared granular system using discete element method simulations. The microslips, i.e., local avalanche events, are found to demonstrate significantly different statistical and spatial characteristics between the stick and slip states. We further investigate the correlation between the global stress fluctuations and the features extracted from microslips based on the machine learning (ML) approach. The data-driven model that incorporates the information of the spatial distribution of microslips can robustly predict the magnitude of stress fluctuation. A further feature importance analysis confirms that the spatial patterns of microslips manifest key information governing the global stress fluctuations.