Fractures in non-porous, subsurface rock is primarily responsible for the flow and transportation of fluids through the medium. The uncertainty in subsurface modeling can require a large ensemble of simulation results to capture the range of behavior of the system, especially when considering large-scale domains. We use a Discrete Fracture Network (DFN) approach for modeling fractured rock, however generating full fidelity flow and transport simulations on DFN is computationally expensive. A recent solution has been to generate realizations using a graph-based emulator trained on a limited set of high-fidelity simulations. The cost of producing results with an emulator is orders of magnitude less than using a full simulation, which enables the production of large ensembles necessary to make predictions. However, the results produced by emulation have greater uncertainty and are more likely to produce non-physical behavior. We have developed a set of feature-based analysis tools for evaluating ensembles produced through emulation. Using clustering and classification methods, we evaluate an ensemble of emulated results by comparing individual realizations to a smaller ensemble produced through high-fidelity simulation.
