Fast prediction of immiscible two-phase displacements in heterogeneous porous media with convolutional neural network

Abstract

A convolutional neural network is developed for rapidly predicting multiphase flow in heterogeneous porous media. Some direct numerical methods can acquire accurate results of multiphase flow in porous media. However, once the geometry of the porous media changes, it takes much computational time to perform a new simulation. Here, a deep neural network model in the field of semantic segmentation is developed. It takes the two-dimensional microstructure of heterogeneous porous media as inputs and is able to predict corresponding multiphase flow fields (pressure and saturation fields). Compared to the direct lattice Boltzmann simulations, the inference time on new geometry of porous media can be reduced by several orders of magnitude. Our results show that the machine learning method is a good prediction tool in a wide range of porosity and heterogeneity. Besides, to better understand the inherent process, a visible explanation is presented on what our neural networks have learned.