François Renard - Can earthquakes be predicted?: predicting the proximity to system-size failure in rocks using X-rays and machine learning

Earthquake, landslide, and glacier instabilities are multi-scale processes that cannot be predicted. 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 catastrophically. Despite extensive research on rock damage, researchers are still far from being able to predict when a rock will fail. Here, I will show how to use a multi-view convolutional neural network model to identify characteristics of a failing rock. By training a neural network model on images of rock samples exposed to different levels of stress, one can predict how close a rock is to failure. The neural network model outperforms traditional estimates based on fracture density. More importantly, the model provides fundamental insight that may offer precursory information on impending material failure. For instance, according to the model, the angle of the fracture plane relative to the principal loading direction becomes a key factor contributing to failure.