Pore-scale modeling of multiphase flow in porous media using a conditional generative adversarial network (cGAN)

Multiphase flow in porous media is involved in various natural and industrial applications, including water infiltration into soils, carbon geosequestration, and underground hydrogen storage. Understanding the invasion morphology at the pore scale is critical for better prediction of flow properties at the continuum scale in partially saturated permeable media. The deep learning method, as a promising technique to estimate the flow transport processes in porous media, has gained significant attention. However, existing works have mainly focused on single-phase flow, whereas the capability of data-driven techniques has yet to be applied to the pore-scale modeling of fluid-fluid displacement in porous media. Here, the conditional generative adversarial network is applied for pore-scale modeling of multiphase flow in two-dimensional porous media. The network is trained based on a data set of porous media generated using a particle-deposition method, with the corresponding invasion morphologies after the displacement processes calculated using a recently developed interface tracking algorithm. The results demonstrate the capability of data-driven techniques in predicting both fluid saturation and spatial distribution. It is also shown that the method can be generalized to estimate fluid distribution under different wetting conditions and particle shapes. This work represents the first effort at the application of the deep learning method for pore-scale modeling of immiscible fluid displacement and highlights the strength of data-driven techniques for surrogate modeling of multiphase flow in porous media.