Volume 49

Rapid Prediction of CO2 Flooding Attribute Distribution in 3D Heterogeneous Reservoirs Using Fourier Neural Operator Zepeng Yang, Xinwei Liao, Peng Dong, Lingfeng Zhang

Abstract

During the implementation of CCUS-EOR projects, clearly defining the distribution of key attributes between injection and production wells relies on reservoir numerical simulation. However, the complex solution process of compositional models results in high computational costs for traditional numerical simulators. Deep learning surrogate models can serve as a reliable alternative to reservoir numerical simulators, significantly improving computational efficiency. This study establishes a surrogate model based on the Fourier Neural Operator (FNO) for 3D heterogeneous CO2 displacement numerical simulation. The model predicts the distribution of pressure, CO2 molar fraction, and oil saturation at each time step using heterogeneous porosity, permeability fields, and injection-production parameters. The research results demonstrate that the developed surrogate model can quickly and accurately predict the distribution of various attributes in the heterogeneous 3D reservoir. It can accurately capture the different displacement characteristics in the miscible and near-miscible regions during the CO2 flooding process. Additionally, the model is able to learn from the data the differences in gas influx across perforated layers in the vertically positive rhythm and reverse rhythmic heterogeneous reservoirs, as well as the impact of gravity override on CO2 displacement characteristics. After training, the surrogate model can achieve a 360-fold improvement in computational efficiency compared to the numerical simulator. The work in this paper has certain application prospects for engineering tasks such as rapid site selection of CO2 injection pilot areas in heterogeneous 3D reservoirs, optimization of injection and production parameters, and determination of the migration direction of the CO2 gas injection front.

Keywords CO2 flooding, numerical simulation, surrogate model, Fourier neural operator, deep learning