Deep-Learning Analysis of Fracture Networks Leading to System-Size Failure in Rocks
Résumé
Fracture networks in rocks and other geomaterials form by seismic and aseismic damage processes that could lead to system-size failure. Here, we use a multi-view convolutional neural network model to predict the stress proximity to macroscopic failure of experimentally deformed rock samples using two-dimensional images. Models are trained on time series of fractures observed in rock samples through dynamic in situ synchrotron X-ray tomography experiments. The results demonstrate that deep learning models outperform traditional estimates based on fracture density, increasing the accuracy of the predictions. Furthermore, the models provide insights into fundamental characteristics of fracture patterns that may provide precursory information on impending material failure. The trained deep learning models estimate the angle of the fracture plane relative to the principal loading direction, which is a key factor contributing to shear failure. The predicted angle of the fracture plane, in the range of 10°–30° with respect to the direction of maximum compressive stress, is consistent with the established failure criteria used in rock mechanics. , Plain Language Summary When a rock is subjected to large stress, it can develop fractures that grow and connect to form a network. This network of fractures spans the entire rock right before it fails. Despite extensive research on rock damage, we are still far from being able to predict when a rock will fail. In our manuscript, we use a multi-view convolutional neural network model to identify characteristics of a failing rock. We train our neural network model on images of rock samples exposed to different levels of stress and use it to predict how close a rock is to failure. Our neural network model outperforms traditional estimates based on fracture density. More importantly, our model provides fundamental insight that may offer precursory information on the imminent material failure. For instance, according to our model, the angle of the fracture plane relative to the principal loading direction becomes a key factor contributing to failure. , Key Points Convolutional neural network models predict the proximity of rock samples to failure The proximity to failure is predicted better with images than with only fracture density Deep learning results find optimal fault angles in the range 10°–30° with respect to σ 1 , in agreement with empirical failure criteria
Domaines
| Origine | Fichiers éditeurs autorisés sur une archive ouverte |
|---|---|
| Licence |
![]()
Cite 10.11582/2019.00001 Jeu de données Renard, F. (2019). Dynamics of microscale precursors establish brittle compressive failure in Carrara marble as a critical phenomenon [Dataset]. Archive2014. https://doi.org/10.11582/2019.00001
![]()
Cite 10.11582/2021.00002 Jeu de données Renard, F. (2021). X-ray tomography data of Westerley granite [Dataset]. Archive2014. https://doi.org/10.11582/2021.00002