Data-Driven Machine Learning Approach to Predict the Trapping Indices in Coupled Chemo-Hydro-Mechanical Model of CO2 Sequestration into Deep Saline Aquifer

Abstract

Abstract Storing carbon dioxide (CO2) in deep geological formations, such as saline aquifers and depleted oil and gas reservoirs, through Geological Carbon Sequestration (GCS) offers tremendous potential for large-scale CO2 storage. However, ensuring the successful implementation of GCS requires a thorough evaluation of the risks associated with confinement of plumes and storage capacity at each storage location. To gain a better understanding of how CO2 is trapped in saline aquifers, it is important to create robust and speedy tools for assessing CO2 trapping efficiency. Therefore, this study focuses on using machine learning techniques to predict the efficiency of CO2 trapping in deep saline formations as part of Geological Carbon Sequestration (GCS). The methodology involves simulating the CO2 trapping mechanisms using a physics-based numerical reservoir simulator and creating training, testing, and validation datasets based on uncertainty variables. The study used a numerical reservoir simulator to simulate CO2 trapping mechanisms over 170 years, with uncertainty variables like petrophysical properties, reservoir physical parameters, and operational decision parameters being utilized to create a large dataset for training, testing, and validation. The study identified key control variables through feature importance index calculation and utilized the Latin-Hypercube approach to account for a wide range of parameters. 722 reservoir simulations were performed and the results of residual trapping, mineral trapping, solubility trapping, and cumulative CO2 injection were analyzed. The outliers and extreme data points were removed using statistical and exploratory data analysis techniques. Deep neural network was applied to predict the CO2 trapping efficiency. The results showed that the deep neural network model can predict the trapping indices with a coefficient of determination above 0.95 and average absolute percentage error below 5%. These findings suggest that machine learning models can serve as a more efficient alternative to traditional numerical simulation for estimating the performance of CO2 trapping in GCS projects.