Prediction of minimum miscibility pressure (MMP) of the crude oil-CO2 systems within a unified and consistent machine learning framework

Abstract

In this study, considering the differences of minimum miscibility pressure (MMP) measured with the slim-tube and rising bubble apparatus (RBA) methods and taking each individual compound and the parameter representing the source of MMP measurement as the input, a new machine learning model has been proposed and validated to determine MMPs of the crude oil-CO2 systems by use of the one-dimensional convolutional neural network (1D-CNN). Then, the differences of the aforementioned two MMP measurement methods and each individual factor on the predicted MMP are analyzed by using the SHapley Additive exPlanations (SHAP). Compared to the existing models that ignore the differences of MMPs obtained from different experiments and taking the pseudocomponents as the input, the newly proposed model not only has the lowest mean absolute percentage error (MAPE) of 7.29%, lowest mean square error (MSE) of 3.7916, and highest R2 of 0.9458 on the slim-tube test dataset, but also can be used to accurately reflect the relationship between the MMP and each of its influential factor. Furthermore, it is observed that there is a very good correlation between the MMP measured with the slim-tube and the RBA methods to capture the corresponding physics in such two experiments within a unified, consistent, and efficient framework.