Interpretable knowledge-guided framework for modeling minimum miscible pressure of CO2-oil system in CO2-EOR projects

Abstract

Carbon dioxide enhanced oil recovery (CO2-EOR) is a promising application for carbon capture, utilization and storage (CCUS). Accurate modeling of CO2-oil minimum miscible pressure (MMP) is crucial for CO2-EOR projects. In this study, a knowledge-guided framework for an extreme gradient boosting machine (XGBoost) and interpretable tabular learning architecture (TabNet), called KXGB and KTabNet, respectively, are developed to model the MMP. The proposed models are strengthened using a large MMP database of 421 samples collected from literature. Domain knowledge is integrated into intelligent models to prevent data-driven models from producing predictions that violated the domain knowledge. The Shapley Additive Explanations (SHAP) method is used to explain the proposed model to ensure the credibility of petroleum engineers. To further verify the model’s effectiveness, the same experimental strategy is employed to compare the proposed models with existing machine-learning (ML) methods. The results show that KXGB is the most recommended superior solution for modeling MMP owing to its outstanding performance and simplicity of optimization. The correlation coefficient, root mean square error, and mean absolute error are 0.9833, 0.7637 and 0.55, respectively. However, KTabNet has great potential. Although its accuracy is slightly lower than that of the former, its strong representation ability and decision transparency may be favorable for future research. This study also demonstrate that the proposed framework conforms to certain theoretical rules and has a reasonable domain of applicability. To the best of our knowledge, this is the first study on integration of domain knowledge into MMP modeling methods. The experience and insights obtained from this study can guide CO2-EOR projects and other tabular data modeling in the oil and gas industry.