GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 1.1 GNGTS 2023 Using AE based Machine Learning Approaches to Forecast Rupture during Rock Deformation Laboratory Experiments S.C. Vinciguerra 1 , T. King 2 , G.M. Adinolfi 1 , Philip Benson 3 1 Department of Earth Sciences, University of Turin, Italy 2 National Buried Infrastructure Facility, University of Birmingham 3 Rock Mechanics Laboratory; School of Earth and Environmental Sciences, University of Portsmouth, United Kingdom The ability to detect precursory activity before dynamic failure of rock samples has key implications for hazard forecasting. Acoustic Emissions (AE) are already a proven analogy to field-scale earthquakes and therefore provide an ideal window into understanding these processes. However, challenges introduced by noise, scale-dependent effects and under- sampling of data means that it is difficult to identify what are the key parameters that we should focus on when upscaling laboratory information to the field. Time Delay Neural Networks (TDNN) have shown promise in forecasting failure when using AE-derived parameters. Still, as with many machine learning techniques, these are "black box" tools and it is not always clear as to why they are working and what we need to do to improve our methods. Here, we expand upon this to identify what are our most important parameters, and more interestingly, when they are most relevant for predicting when the rock will fail. A conventional triaxial apparatus (Sanchez Technologies) was used for experimentation. The apparatus is designed to deform samples of 40 D and 100 mm L specimens at confining pressures of up to 100 MPa at strain rates of 3.6 mm/hr. Four samples of Alzo Granite were deformed at 5, 10, 20 and 40 MPa (Fig. 1). During deformation, AE are recorded by an array of 12 1MHz