Using AE based Machine Learning approaches to forecast dynamic failure during rock deformation laboratory experiments.
Abstract
The ability to predict dynamic failure of brittle rock media would usher in a new era of geophysics and hazard forecasting of crustal scale processes. In recent years, laboratory scale rock deformation experiments are providing a wealth of information on the physics of the fracture process ranging from fracture nucleation, crack growth and damage accumulation, to crack coalescence and strain localisation. Parametric analysis of Acoustic Emission (AE) data has revealed periodic trends and precursory behaviour of the rupture source mechanisms as a fault zone enucleates and develops, suggesting these processes are somehow repeatable and forecastable. However, due to the inherent anisotropy of rock media and the range of environmental conditions in which deformation occurs, finding full consistency between AE datasets and a prediction of rupture mechanisms from AE analysis is still an open goal. Nevertheless, recent advances in artificial intelligence has led to the development of a suite of user-friendly tools that may provide a way forward. A promising direction is in the form of black-box non-linear autoregressive network models with exogenous inputs (NARX). In these models, the next value of a dependent output signal is predicted using previous values of the output as well as previous values of additional independent (exogenous) parameters. In this study, we apply a NARX network to AE source parameters (i.e. focal mechanism, amplitude, AE rate) for forecasting thresholds of strain in triaxially deformed at crustal effective pressure of different lithologies (e.g. granites and sandstones) brought to failure. Although a relatively high number of input events were required to forecast any number of timesteps into the future, the strain rate was effectively reproduced with an accurate prediction of when sample failure would occur. We expanded on this with a Markov-switching dynamic regression models. These work similarly to NARX forecasting but are able to maintain multiple models which can characterize individual time periods, as damage develops during rock deformation experiments. A transition matrix then determines when and if one model may switch to another, forecasting the mechanical transition to a different deformation patterns such crack closure, growth, coalescence and strain localization.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMMR0100012V
- Keywords:
-
- 1236 Rheology of the lithosphere and mantle;
- GEODESY AND GRAVITY;
- 5104 Fracture and flow;
- PHYSICAL PROPERTIES OF ROCKS;
- 7209 Earthquake dynamics;
- SEISMOLOGY;
- 8163 Rheology and friction of fault zones;
- TECTONOPHYSICS