Simulating multiphase flow in fractured media with graph neural networks

Abstract

Numerical modeling of flow dynamics with multiple fluid phases in subsurface fractured porous media is of great significance to numerous geoscience applications. Discrete fracture-matrix (DFM) approaches become popular for simulating fractured reservoirs in the last decade. Data-driven surrogate models can provide computationally efficient alternatives to high-fidelity numerical simulators. Although convolutional neural networks (CNNs) are effective at approximating the space-time solutions of multiphase flowing processes, it remains difficult for CNNs to operate upon DFMs with unstructured meshes. To tackle this challenge, we leverage graph neural networks (GNNs) for surrogate modeling of an embedded DFM model. The results of two-dimensional cases with complex fracture systems show that the learned surrogates precisely capture the effect of the variations in fracture connectivity and forecast dynamic pressure and saturation solutions with high accuracy. Furthermore, our GNN-based models exhibit promising generalizability to fracture networks with different geometries and numbers of fractures that are not encountered from the training dataset.