Discrete Fracture Network insights by eXplainable AI

In the framework of flow simulations in Discrete Fracture Networks (DFN), we consider the problem of identifying backbones. Backbones of a DFN are subnetworks of fractures that preserve the main flow properties and can be fruitfully used to reduce the computational cost of simulations or to analyze clogging and waste storage problems. With a well-trained Neural Network for flux regression in a DFN at hand, we use the Layer-wise Relevance Propagation method to compute the expected relevance of each fracture to identify the backbone.