Optimizing CO2 Sequestration in Coal Seams: A Machine Learning Framework for Wettability Prediction

Abstract In the face of escalating global climate challenges, Carbon Capture, Utilization, and Storage (CCUS) emerges as a pivotal technology in the quest to mitigate atmospheric CO2 emissions. Among the various geological formations suitable for CO2 sequestration, coal seams represent a unique and promising option, especially those that are unmineable due to depth, thickness, or other economic constraints. The success of this formation is closely tied to the intricate interactions among CO2, coal, and in-situ fluids, with wettability playing a crucial role. A deep understanding of wettability is pivotal for optimizing CO2 injection strategies and ensuring the long-term stability of the sequestered CO2. Over the past decade, researchers have increasingly turned to data-driven methods to predict rock-fluid interactions, yet the limited size of datasets has constrained the representativeness and applicability of their results. Additionally, traditional methods for assessing wettability, such as contact angle measurements, Amott test and interpolation techniques involving nuclear magnetic resonance, are costly and time-consuming. In response to these challenges, this study employs a data-driven approach, leveraging a collection of experimental datasets to predict wettability in a coal/CO2/brine system. The framework incorporates features such as physical properties of coal, ambient conditions, CO2 characteristics, coal rank, and surface chemistry, utilizing advanced data analysis techniques such as heatmaps, cross-validation, feature engineering and importance analysis to enhance model generalizability. Four machine learning models were employed in this study: Bayesian Linear Regression, Explainable Boosting Machines (EBMs), Google's TabNet algorithm and a composite of the last two models (Voting Regressor.) The voting regressor model demonstrated superior predictability with a coefficient of determination (R2) score of 0.86, mean absolute percentage error of 5.32% and root mean squared error of 5.92 on the blind test set, outperforming the other stand-alone models. The strong correlation coefficient of 0.954 between measured and predicted wettability values underscores the model's robustness. Consequently, this study advances the predictability of key parameters for CO2 sequestration and underscores the feasibility of using coal seams for long-term CO2 storage, significantly contributing to CCUS research.