Interpretable Machine Learning for Prediction of Minimum Miscibility Pressure in CO2-Oil System Considering Nano-Confinement Effect

Abstract

Abstract The determination of the minimum miscibility pressure (MMP) in CO2-oil systems is critical for modeling CO2-EOR processes experimentally and numerically. Nevertheless, in nano-confined space, the existing experimental and empirical formula methods present limitations regarding the utilization conditions and prediction accuracy respectively. Thus, in this study, a novel approach combining ML model with Shapley Additive Explanations (SHAP) algorithm is introduced, which aims to provide more precise and physically correct estimates of the MMPs considering the influence of nano-confinement. A database containing MMPs in CO2 injection process under different conditions is firstly established based on 348 samples collected from experimental results and open publications. The input parameters determining MMPs include reservoir temperature, pore size, and oil composition. In this framework, XGBoost and MLP are used to mimic the input-output relations of the database. Then, SHAP is employed to comprehensively interpret the impact of the inputting factors on the MMPs by calculating the SHAP values. The present study revealed that both the proposed XGBoost and MLP models exhibited R2 score exceeding 80% and demonstrated good predictive accuracy, as evidenced by small MAE, MSE, and MAPE values. Moreover, a comparative analysis of the SHAP interpretation results of the two models revealed that the explanatory patterns of the MLP model were more consistent with established physical laws, thereby rendering it more suitable for constructing an MMP prediction model based on the dataset employed in this investigation. It is noteworthy that although the SHAP interpretation of the XGBoost model did not entirely conform to actual physical laws, the influence of pore size on MMP followed the same pattern as elucidated by the MLP model. Specifically, within the nano-confined spaces, MMP decreased as the pore size decreased, and the pore size played a crucial role in predicting MMP (ranking first in the XGBoost model and second in the MLP model). The outcomes demonstrate that the developed interpretable machine learning framework, which incorporates the effects of nano-confinement, can accurately predicts MMP under diverse conditions while maintaining the consistency of physical laws. Consequently, this framework offers valuable insights for the implementation and optimization of CO2-enhanced oil recovery processes.