Multiscale numerical simulation of CO2 hydrate storage using machine learning

Carbon capture and storage (CCS) is a technology to reduce significant emissions of carbon dioxide (CO2). CO2 can be stored in sub-seabed aquifers under impermeable layers called caprock. CO2 hydrate forms under the conditions of high pressure and low temperature and it may suppress or block CO2 upwards flow as functioned as a kind of 'caprock'. Therefore, understanding of permeability reduction mechanisms due to hydrate formation is essential for CCS using gas hydrate. This study proposes a novel multiscale numerical method, by which the permeability of hydrate-bearing sediments in reservoir-scale simulations of several hundred meters is estimated using a machine learning technique based on the database of the results of microscopic pore-scale simulations of a couple of hundred μm. Because the permeability reduction is strongly influenced by the characteristics of microscale hydrate distribution, we used two types of neural networks: one is for parameters with regard to the hydrate shape distributed in the pore space and the other is for the permeability reduction due to the hydrate shape. Reservoir-scale simulations were conducted in homogeneous sand sediments using the proposed machine learning model, based on the database of the microscale simulation results, together with existing mathematical permeability models for comparison. The results of the reservoir-scale hydrate distribution using these models were very similar, indicating that the proposed method was validated, while at the same time the difference in the results of the spatial distribution of permeability values, albeit small, suggests the originality of this method.