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

摘要: Structural and topological information play a key role in modeling flow transport through fractured rock the subsurface. Discrete fracture network (DFN) computational suites such as dfnWorks are designed to simulate porous media. Flow calculations reveal that small backbone of fractures exists, where most occurs. Restricting flowing this provides significant reduction network's effective size. However, particle tracking simulations needed determine computationally intensive. Such methods may be impractical for large systems or robust uncertainty quantification networks, thousands forward bound system behavior. In paper, we develop an alternative approach characterizing DFNs, by combining graph theoretical machine learning methods. We consider representation nodes signify edges denote their intersections. Using random forest support vector machines, rapidly identify subnetwork captures patterns full DFN, based primarily on node centrality features graph. Our supervised techniques train particle-tracking paths found dfnWorks, but run negligible time compared those simulations. find our predictions can reduce approximately 20% its original size, while still generating breakthrough curves consistent with network.