A deep learning-accelerated data assimilation and forecasting workflow for commercial-scale geologic carbon storage

Fast assimilation of monitoring data to update forecasts pressure buildup and carbon dioxide (CO2) plume migration under geologic uncertainties is a challenging problem in storage. The high computational cost with high-dimensional parameter space impedes fast decision-making for commercial-scale reservoir management. We propose leverage physical understandings porous medium flow behavior deep learning techniques develop assimilation-reservoir response forecasting workflow. Applying an Ensemble Smoother Multiple Data Assimilation (ES-MDA) framework, the workflow updates properties predicts performance quantified uncertainty from history CO2 plumes interpreted through seismic inversion. As most computationally expensive component such simulation, we developed surrogate models predict dynamic extents multi-well injection. employ convolutional neural networks, specifically, wide residual network U-Net. validated against flat three-dimensional model representative clastic shelf depositional environment. Intelligent treatments are applied bridge between quantities true-3D those single-layer model. can complete matching quantification less than one hour on mainstream personal workstation.