Intelligent Approach for Estimating CO2-Brine Relative Permeability in Carbon Capture and Storage Process Based on Deep Learning: A Case Study of Aquifer Under the South China Sea

Abstract Geological storage of carbon dioxide (CO2) in the brine aquifer is a dual-benefit strategy for mitigating greenhouse gases and advancing carbon capture and storage (CCS). Meanwhile, the gas-water relative permeability (RP) of CO2 and brine provides fundamental insights into flow behavior in the aquifer which plays an essential role in evaluating the carbon storage capacity. Consequently, establishing an economic and useful approach for obtaining the gas-water RP curve in the CCS process is urgently needed. In this study, an intelligent yet accurate model, called DL-MICP, for estimating the gas-water RP curve in the CCS process adopting deep learning (DL) techniques and experimental mercury injection capillary pressure (MICP). Specifically, the DL methods and experimental data including the Gramian angular field (GAF), the convolutional long short-term memory (ConvLSTM), self-supervised learning framework, curves of mercury injection and withdrawal, and distribution of pore size are utilized in the DL-MICP approach for evaluating CO2-brine RP curves. To validate the superior and effective performance of the DL-MICP approach, the experimental data of aquifer samples under the South China sea is applied, in which 38 small-sample data spaces are utilized for training and 9 samples for testing. It is found that the average errors are 3.32% and 2.58% for the CO2 and brine RP curves, respectively. In addition, the superior accuracy of the DL-MICP method is demonstrated through the comparative study against classical convolutional neural network (CNN) and long short-term memory (LSTM) models. This study combined deep learning techniques and mercury injection capillary pressure to efficiently realize the rapid intelligent prediction of CO2-brine RP curves that facilitate the evaluation of the CO2 injection capacity and flow behaviors in the CCS process. To this end, this intelligent DL-MICP method achieved the goal of obtaining CO2-brine RP curves in excellent accordance with experimental data without costly and time-consuming experimental measuring.