The U-Net Enhanced Graph Neural Network for Multiphase Flow Prediction: An Implication to Geological Carbon Sequestration

Abstract Monitoring CO2 pressure buildup and saturation plume movement throughout the operation of Geological Carbon Sequestration (GCS) projects is crucial for ensuring environmental safety. While the movement of CO2 plumes can be predicted with high-fidelity numerical simulations, these simulations are often computationally expensive. However, through training on readily available simulation datasets, recent advancements in data-driven models have made it possible to predict CO2 movement rapidly. In this study, we adopt the U-Net Enhanced Graph Convolutional Neural Network (U-GCN) to predict the spatial and temporal evolution of CO2 plume saturation and pressure buildup in a saline aquifer reservoir. Utilizing the U-Net architecture, which incorporates skip connections, enables U-GCN to capture high-level features and fine-grained details concurrently. First, we construct physics-based numerical simulation models that account for both GCS injection and post-injection periods. By employing Latin-Hypercube sampling, we generate a diverse range of reservoir and decision parameters, resulting in a comprehensive simulation database comprising 2000 simulation cases. We train and test the U-GCN model on a two-dimensional (2D) radial model to establish a U-GCN code benchmark. We utilize Mean Squared Error as the loss function throughout the U-GCN training process. The U-GCN model demonstrates robust performance on the radial model, achieving an R2 score of 0.993 for saturation prediction and an R2 of 0.989 for pressure buildup prediction based on the blind testing dataset. Notably, the Mean Absolute Percentage Error (MAPE) for all mappings consistently hovers around less than 5%, indicating the effectiveness of the trained models in predicting the temporal and spatial evolution of CO2 gas saturation. Moreover, the prediction CPU time for the DL models is significantly lower (0.02 seconds per case) than the physics-based reservoir simulator (on average, 45 to 60 minutes per case). This underscores the capability of the proposed method to provide predictions as accurate as physics-based simulations while reducing substantial computational costs.