TransUNet: Accelerating Multiphase Flow Prediction in Geological Carbon Storage Systems

Abstract Deep learning-based surrogate models provide an effective alternative to numerical simulations for addressing subsurface multiphase flow challenges, such as those encountered in Geological Carbon Storage (GCS). In this study, we implemented deep learning based surrogate model as an alternative to complex GCS simulations using TransUNet, an enhanced U-Net architecture that incorporates Transformer models. TransUNet predicts the spatial and temporal evolution of CO2 plume saturation and pressure buildup in saline aquifers by leveraging the capabilities of Transformers. TransUNet is designed to effectively extract features from the structured data, considering spatial relationships and leveraging the Transformer architecture to capture both high-level and detailed information concurrently. Initially, we established physics-based numerical simulation models to account for both injection and post-injection periods of GCS. Employing Latin-Hypercube sampling, we generate a diverse range of reservoir and decision parameters, resulting in a comprehensive simulation databases. We train and test the TransUNet model on two different datasets: a radial model to establish a code benchmark, and the 2D complex model to validate the performance efficiency. Throughout the TransUNet training process, we utilize Mean Squared Error and the spatial derivative as the loss functions. The TransUNet model demonstrates robust performance on the radial model, achieving an R2 of 0.9982 and 0.9963 on testing dataset for saturation and pressure buildup predictions, respectively. Similarly, the model with updated hyper-parameters exhibits comparable performance on the 2D complex model, with R2 values of 0.9986 and 0.9967 on testing dataset for saturation and pressure buildup predictions, respectively. Notably, the Normalized Absolute Error (NAE) for all mappings consistently hovers around 1%, 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 TransUNet model is significantly lower at 0.02 seconds per case compared to the physics-based reservoir simulator's, 2500 seconds per case for the radial model and 1500 seconds for the 2D complex Cartesian model. This underscores the capability of the proposed method to provide predictions as accurate as physics-based simulations while offering substantial computational timesavings.