Session: 11-11-01 Advances in Carbon Capture Utilization and Storage

Paper Number: 78523

78523 - Development of a Permeability Reduction Model Using Deep Learning for CO2 Hydrate Storage 

   Global warming is one of the biggest environmental concerns and the necessity to reduce the emission of greenhouse gases is very important. Carbon capture and storage (CCS) emerges as an important and promising process for this task whereas carbon dioxide is captured from large emission sources and it is later transported and stored under aquifers onshore or seabed regions offshore. The carbon dioxide can be stored under layers which have very low permeability called cap-rock. A risk of leakage exists due to possible fractures in the cap-rock, although such a risk is very small. There is a leak-trapping mechanism that utilizes the natural formation of carbon dioxide (CO₂) hydrate, which can create a new low-permeability layer. This concept can also be extended to be used by injecting CO₂ directly in the hydrate stability zone, i.e. regions with proper conditions of pressure and temperature for its formation, allowing for the formation of a low-permeability layer of gas hydrate. 

   Numerical simulations are used to study how the CO₂ hydrate formation affects the permeability change in porous media. The main parameter in this study is the permeability reduction coefficient defined as the ratio between the permeability after hydrate formation and the original permeability. Various mathematical permeability models have already been proposed relying mostly on empirical data for different sand formations. In these models, the main parameter consists mainly of the hydrate saturation and an empirical number that represents the change in porous shape due to hydrate distribution within the pore space. In this study, a new permeability reduction model is proposed: that is, a multiscale approach using deep learning techniques.

   For the large-scale simulation, a commercial program, General-purpose Terrestrial Fluid-Flow Simulator (GETFLOWS), is used to simulate the two-phase flow and hydrate formation, which is mathematically modelled considering that hydrate is formed in three different regions: on the CO₂ front, on the hydrate film behind the CO₂ front and on the surface of the sand grains behind the CO₂ front. On the other hand, hydrate formation in the microscale simulation is analyzed using the Phase Field model, which can offer the hydrate shape information within the pore space, such as the tortuosity and surface area of CO₂ hydrate.

   However, the coupling of these two different scales is very difficult because the computational cost will be enormous when the microscale simulations are called at every time step and at every computational cell of the large-scale simulation. To make such a coupling realistic and effective, a deep learning method was applied by learning tons of results of the microscale simulations and giving a permeability value at each cell of the large-scale simulation at each time step. Two neural network models were used for this purpose: the first one obtains information regarding the hydrate shape by predicting how it changes with time; the second gives the new permeability reduction coefficient based on the parameters predicted in the first neural network. The validation was conducted by comparing the large-scale simulation results using the proposed method and a conventional mathematical model for permeability reduction.

Presenting Author: Alan Junji Yamaguchi The University of Tokyo

Authors:

Alan Junji Yamaguchi The University of Tokyo
Toru Sato The University of Tokyo
Takaomi Tobase Electric Power Development Co. Ltd.
Xinran Wei Microsoft Research Asia
Lin Huang Microsoft Research Asia
Jia Zhang Microsoft Research Asia
Jiang Bian Microsoft Research Asia
Tie-Yan Liu Microsoft Research Asia