Simulation Enhancement GAN for Efficient Reservoir Simulation at Fine Scales

Abstract

In this paper, an innovative approach for enhancing fluid transport modeling in porous media is presented, which finds application in various fields, including subsurface reservoir modeling. Fluid flow models are typically solved numerically by addressing a system of partial differential equations (PDEs) using methods such as finite difference and finite volume. However, these processes can be computationally demanding, particularly when aiming for high precision on a fine scale. Researchers have increasingly turned to machine learning to explore solutions for PDEs in order to improve simulation efficiency. The proposed method combines an adaptive multi-scale strategy with generative adversarial networks (GAN) to increase simulation efficiency on a fine scale. The devised model, called simulation enhancement GAN (SE-GAN), takes coarse-scale simulation results as input and generates fine-scale results in conjunction with the provided petrophysical properties. With this new approach, a deep learning model is trained to map coarse-scale results to fine-scale outcomes, rather than directly solving the fluid flow model. Case studies reveal that SE-GAN can achieve a significant improvement in accuracy while reducing computational time compared to the original fine-scale simulation solver. A comprehensive evaluation of numerical experiments is conducted to elucidate the benefits and limitations of this method. The potential of SE-GAN in accelerating the numerical solver for reservoir simulations is also demonstrated.