Deep-Learning-Based Flow Prediction for CO2 Storage in Shale–Sandstone Formations

Abstract

Carbon capture and storage (CCS) is an essential technology for achieving carbon neutrality. Depositional environments with sandstone and interbedded shale layers are promising for CO2 storage because they can retain CO2 beneath continuous and discontinuous shale layers. However, conventional numerical simulation of shale–sandstone systems is computationally challenging due to the large contrast in properties between the shale and sandstone layers and significant impact of thin shale layers on CO2 migration. Extending recent advancements in Fourier neural operators (FNOs), we propose a new deep learning architecture, the RU-FNO, to predict CO2 migration in complex shale–sandstone reservoirs under various reservoir conditions, injection designs, and rock properties. The gas saturation plume and pressure buildup predictions of the RU-FNO model are 8000-times faster than traditional numerical models and exhibit remarkable accuracy. We utilize the model’s fast prediction to investigate the impact of shale layer characteristics on plume migration and pressure buildup. These case studies show that shale–sandstone reservoirs with moderate heterogeneity and spatial continuity can minimize the plume footprint and maximize storage efficiency.

Machine learning for underground gas storage with cushion CO2 using data from reservoir simulation

“…For example, with the addition of well-block pressure and average reservoir pressure, the ANN model could contribute to establishing their relations to well deliverability [7,8]. It is also desirable to explore deeplearning methods to describe the CH 4 /CO 2 distribution in the reservoir, in particular the extent of mixing zones, potentially by utilizing methods developed for CO 2 storage that describe the migration of CO 2 plumes in sandstone formations [6].…”

“…In the literature, reservoir simulation studies of UNGS with cushion gas focus on demonstrating the advantageous density changes of CO 2 near its critical point in idealized reservoirs [3]; the mechanical responses to pressure build-ups in idealized, yet inclined reservoirs IOP Publishing doi:10.1088/1757-899X/1294/1/012058 2 [4]; and simulations of working gas quality caused by mixing with cushion gas [5]. While the entry of data-driven methods has proven useful in research on CO 2 storage [6], there are to date few data-driven methodologies demonstrated for UNGS with cushion CO 2 in reservoirs or aquifers. In the literature on UNGS, most data-driven approaches are concerned with describing the relations between gas deliverability (that is, the amount of gas that can be withdrawn per day [2]), the flowing bottom hole pressure (BHP) in the well, and reservoir pressure [7,8].…”

Prediction of Optimal Production Time during Underground CH4 Storage with Cushion CO2 Using Reservoir Simulations and Artificial Neural Networks

“…The entry of data-driven methods has proven useful in research on CO 2 storage − and on CO 2 injection for enhanced oil recovery, but there are to date few data-driven methodologies demonstrated for UNGS in depleted gas reservoirs or aquifers. In the literature, most data-driven approaches are concerned with describing the relations between gas deliverability, bottom hole pressure (BHP) in the well, and reservoir pressure. , On the other hand, Mann and Ayala developed an artificial neural network (ANN) model aimed at determining the optimum design of a storage facility, but their model did not include information on which gas was considered as cushion gas, nor the reservoir conditions, both of which are important for the behavior of the gases and their mixtures.…”

Fast and Robust Prediction of Multiphase Flow in Complex Fractured Reservoir Using a Fourier Neural Operator

“…The data-driven based machine learning (ML) method is a particular type of artificial neural network that has found tremendous strides in subsurface flow prediction. Several machine learning-based surrogates have been proposed to provide runtimes of several orders of magnitude faster and more accurately than numerical simulations [25,26]. Most existing classic data-driven methods are mainly based on convolutional neural networks (CNNs), that concentrate on learning Euclidean space mappings from traditional numerical simulation data [27][28][29].…”