Machine learning of dual porosity model closures from discrete fracture simulations

Fine-scale discrete fracture simulations provide a natural means to quantify the matrix-fracture fluxes and to specify reference solutions for upscaling approaches such as dual porosity/dual permeability models. Since typically the fine-scale simulations are computationally demanding, and the fractured reservoirs are highly heterogeneous, it is desirable to parametrize the fracture geometry and to obtain coarse-scale model closures using precomputed fine-scale results. We show that this can be done for the case of two-dimensional geometries and compressible single-phase flows. Specifically, a set of parameters linked to a coarse-scale grid block can be mapped to the underlying fracture geometry via a convolutional neural network. In particular, if a matrix-fracture transfer function can be parametrized with a number of parameters spatially varying on a coarse scale, the shape of the transfer function per grid block can be learned from fine-scale simulations.