Modeling CO2 Geologic Storage Using Machine Learning

Abstract

Over the upcoming years, storing CO2 into geological formations would contribute significantly to the international efforts to address climate challenges due to greenhouse gas emissions. Carbon Capture and Storage (CCS) projects require immense capital investments to complete multiple phases, namely careful site selection, planning, design, and execution. Modeling of surface and subsurface CO2 flow plays a major role not only in design optimization but also in site screening and capacity estimation. This study focuses on modeling multiphase flow of CO2 in underground formations with particular emphasis on the fraction of the CO2 injected that can be trapped. Key interest is given to a trapping mechanism that can keep CO2 stored for long-term in target formations, namely residual trapping. The main objective of this work is to find more efficient ways to proxy model this process with its complex physics. There have been multiple recent reservoir simulator numerical enhancements to model CO2 trapping in CCS accurately. However, these complex enhancements have created computational difficulties when attempting to capture unique CO2 fluid physical and chemical subsurface processes such as relative permeability hysteresis. Such numerical challenges make the modeling inefficient and computationally expensive. Therefore, this study introduces more-efficient modeling techniques based on machine learning to make simulations more practical and accessible. By generating a sufficiently large training dataset utilizing a computationally-enhanced numerical simulator with a wide range of input parameters including permeability, porosity, etc., a machine learning model was constructed as an alternative to conventional numerical simulation. The effectiveness of the machine-learning models is presented using a test case of a 2D rectangular grid domain of heterogeneous permeability representing a saline aquifer. The goal is to model an injection of CO2 into water under gravity segregation to estimate the fraction of CO2 trapped at the bottom prevented from reaching the top. Even though the training dataset used in this study is relatively small, the machine-learning alternative is able to achieve at least 95% accuracy when tested with new input data in 103 to 104 faster run-times. It is believed that this accuracy can be improved further by increasing the size of the training dataset and exploring other machine-learning models with new hyperparameters. In this study, only a limited number of widely-used techniques is compared, including: Random Forest, K-Nearest Neighbors (KNN), and Multi-Output Regression. Accurately modeling the amount of CO2 that can be trapped during CCS applications is vital as this will dictate the available storage capacity for injected CO2; however, this may be a difficult task for most commercial simulators. This study proposes new ways to model such process with enhanced efficiency compared to existing techniques. As most global efforts are adopting an accelerated strategy towards sustainability, the presented approach is timely and of great importance.