Predicting CO2-EOR and storage in low-permeability reservoirs with deep learning-based surrogate flow models

Addressing the greenhouse effect, enhancing carbon reduction and utilization has become globally imperative. China has ambitious goals to carbon peak by 2030 and achieve carbon neutrality by 2060. One promising method is CO2 enhanced oil recovery (EOR) technology, which bolsters the development of low-permeability oil reservoirs and offers potential for carbon storage. While numerical simulation can optimize production, multiphase flow simulations often come with considerable computational demands. This study introduces a novel workflow to develop a high-fidelity surrogate model tailored for CO2-EOR in low-permeability reservoirs. Adopting a data-driven methodology, we employ the Res U-Net neural architecture to craft this surrogate model, aiming to predict formation flow state variables. Critical parameters such as porosity, permeability, well location, control conditions, and discrete time intervals from raw data are transformed into 2D images, serving as input features for the model. We incorporate a structural similarity (SSIM) index into the loss function to effectively grasp the intricate spatiotemporal changes inherent in stratigraphic state variables. Additionally, a bespoke penalty scheme refines the surrogate model's performance in designated areas. Compared to numerical simulation results, the surrogate model can achieve high-fidelity prediction of CO2-EOR and storage in heterogeneous low-permeability oil reservoirs. In test set cases, the average MSE of the surrogate model for prediction results is less than 2.2 x 10-5, and the average MSSIM is more significant than 0.99. Based on a post-processing block, we have further predicted cumulative oil production and CO2 storage using the surrogate model, with an average relative error of less than 0.90% for predicted results.

Automated summary

This study introduces a novel workflow to develop a high-fidelity surrogate model tailored for CO2 enhanced oil recovery in low-permeability reservoirs. By adopting a data-driven approach and utilizing a neural network, the study improves the accuracy of predicting formation flow state variables and compares the results with numerical simulation. The findings demonstrate that the surrogate model can achieve high-fidelity prediction of CO2 enhanced oil recovery and storage.