Machine learning for graph-based representations of three-dimensional discrete fracture networks

Structural and topological information play a key role in modeling of flow through fractured rock. Discrete fracture network (DFN) computational suites such as dfnWorks [23] are designed to simulate flow and transport in such media. Transport calculations that use a particle tracking method reveal that a small backbone of fractures exists where most transport occurs providing a significant reduction in the effective size of the flowing fracture network. However, the simulations needed for particle tracking are computationally intensive, and may not be scalable to large systems or for robust uncertainty quantification of fracture networks where thousands of forward simulations are needed to bound system behavior. In this paper, we combine machine learning and graph theoretical methods to develop an emulator of dfnWorks for quick estimates of transport that can mimic the high fidelity discrete fracture networks. We introduce a machine learning approach to characterizing transport in DFNs. We consider a graph representation where nodes signify fractures and edges denote their intersections. Using supervised learning techniques, random forest and support vector machines, that train on particle-tracking backbone paths found by dfnWorks, we predict whether or not fractures conduct significant flow, based primarily on node centrality features in the graph. Our methods run in negligible time compared to particle-tracking simulations. We find that our predicted backbone can reduce the network to approximately 20% of its original size, while still generating breakthrough curves resembling those of the full network.