Deep-Learning Framework Forecasts Dynamics of Carbon Dioxide Mineralization

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 216998, “A Deep-Learning Framework To Forecast Spatial-Temporal Dynamics of CO2 Mineralization in Reactive Rocks,” by Zeeshan Tariq, Bicheng Yan, SPE, and Shuyu Sun, SPE, King Abdullah University of Science and Technology. The paper has not been peer reviewed. _ Reactive rocks, when exposed to CO2-charged waters, can undergo a series of reactions leading to the formation of stable carbonates. These carbonates can store carbon for thousands of years. To better understand the interplay between CO2 and brine in these reactive formations, numerical simulations are a useful tool. However, simulating fluid flow in these reservoirs can pose significant computational challenges. The objective of the study detailed in the complete paper is to develop a deep-learning work flow that can predict the changes in CO2 mineralization over time and space in saline aquifers, offering a more-efficient approach compared with traditional physics-based simulations. Methodology Geological Carbon Storage (GCS) Simulator. The study used the Computer Modelling Group’s GEM simulator to simulate the flow of CO2 in a porous media. The numerical simulator is used for conducting multiple simulations and generating data on multiphase flow. The simulation technique used by the simulator is the finite-difference implicit method. The software uses an equation of state to determine fluid properties and uses a range of rock/fluid correlations to estimate properties of the rock, such as capillary pressures and relative permeabilities. Additionally, the simulator can integrate a geomechanical module that considers both plastic and nonplastic strains. The simulator, based on physics principles, implements a finite-difference numerical scheme with upstream weighting for dividing space into discrete gridblocks and uses the adaptive implicit method for dividing time into discrete intervals. Various equations representing mathematical formulations important to the work are provided in the complete paper. UNet Architecture. In this study, the authors used UNet, a deep convolutional neural network, to predict the precipitation of minerals such as siderite and magnesite during CO2 storage in subsurface reactive formation. The architecture has proven effective in various image-processing tasks. Its fully convolutional network structure makes it suitable for image-to-image translation. The network architecture of the UNet consists of multiple layers, each with a distinct type and size. These layers include the input, padding, convolutional unit, max pooling, deconvolution or transposed convolution layers, merging or concatenation layers, and a depadding layer. The convolutional unit performs convolutions using filters of different sizes and strides, whereas the max-pooling layers downsample the feature maps. The deconvolution layers upsample the feature maps, and the merging layers combine the feature maps from different parts of the network.