Deep Learning Accelerated Inverse Modeling and Forecasting for Large-Scale Geologic CO2 Sequestration

Traditional physics-simulation based approaches for inverse modeling and forecasting in geologic CO2 sequestration (GCS) are very time consuming. A single inverse modeling and forecasting process using traditional physics-based approaches (e.g., Ensemble Smoother with Multiple Data Assimilation or Ensemble Kalman Filter) may take a few weeks for a large-scale CO2 storage model (~million grid cells) without leveraging any high-performance computing. To speed up this process, researchers from the U.S. Department of Energy’s SMART Initiative (https://edx.netl.doe.gov/smart/) have developed multiple approaches that employ machine learning methods to integrate monitoring data into subsurface forecasts more rapidly than current physics-based inverse modeling workflows allow. These updated forecasts with the updated models from the inverse modeling (history matching) process will be used to provide site operators with decision support by generating real-time performance metrics of CO2 storage (e.g., CO2 plume and pressure area of review). Here, we present one such machine learning accelerated workflow that can speed up the inverse modeling and forecasting process by three orders of magnitude. First, we developed a deep learning (DL) model to predict the pressure/saturation evolution in large-scale storage reservoirs. A feature coarsening technique was applied to extract the most representative information and perform the training and prediction at the coarse scale, and to further recover the resolution at the fine scale by 2D piecewise cubic interpolation. The accuracy of the feature coarsening-based DL model is validated with a reservoir model (~1.34 million grid cells) built upon a Clastic Shelf storage site. The overall mean relative error between the ground truth and the predictions from DL workflow is no more than 0.2%. Thereafter, the feature coarsening based deep learning model was utilized as forward model in the inverse modeling process where a classical data assimilation approach, ES-MDA-GEO, was applied. The efficiency and effectiveness of the proposed deep learning assisted workflow for large-scale inverse modeling and forecasting was demonstrated with the Clastic Shelf storage model.