Layer-wise relevance propagation for backbone identification in discrete fracture networks

In the framework of flow simulations in Discrete Fracture Networks, we consider the problem of identifying possible backbones, namely preferential channels in the network. Backbones can indeed be fruitfully used to analyze clogging or leakage, relevant for example in waste storage problems, or to reduce the computational cost of simulations. With a suitably trained Neural Network at hand, we use the Layer-wise Relevance Propagation as a feature selection method to detect the expected relevance of each fracture in a Discrete Fracture Network and thus identifying the backbone.