Modeling and simulation of multiphase flow in highly fractured porous media with a data-driven multiscale approach

Abstract The pseudo-direct numerical simulation (P-DNS) method is a recently developed multiscale strategy designed for high-fidelity computational simulation of complex flow physics. This physics-based data-driven approach involves numerically solving both the fine and global scales. The former is precomputed into representative volume elements (RVEs), whose homogenized responses serve to train machine learning-based surrogate models. This upscaling model feeds the global scale, which is then effectively solved in coarse meshes. In this work, the multiscale P-DNS method is applied to the study of multiphase flow in highly fractured porous media. The aim is overcoming the current limitations of simulation techniques for oil reservoirs due to the complex geological heterogeneities. A novel characterization of the geometry of the fracture networks is proposed. The local intrinsic permeability tensor is homogenized via RVE simulations accounting for embedded fractures, thus allowing efficient computation of reservoir-scale transport. The multiscale method is applied to two-dimensional single-phase and two-phase flow problems on different reservoir scenarios. The accuracy of the predictions is assessed relative to detailed simulations with embedded fractures on very fine meshes. For the cases considered, it is shown that the P-DNS homogenization technique is capable of compute accurate flow rates and pressure fields on coarser meshes than the high-fidelity approach, while achieving speedups in the solution time of about a factor of 500.