MACHINE LEARNING ANALYSES OF LOW SALINITY EFFECT IN SANDSTONE POROUS MEDIA
Lei Wang, Ye Tian, Bowen Yao, Xiangyu Yu
Abstract
Multiple enhanced oil recovery (EOR) mechanisms have been proposed for low salinity water flooding (LSWF) in sandstone porous media, but none of them could explain all cases. Previous linear regression analyses showed that multiple conditions must be met to activate intermediate oil-brine-rock interactions to achieve incremental oil recovery, i.e., low salinity effect. However, the linear relationships found are of merely "moderate" strength with high mean squared errors (MSE), which are insufficient for prediction analysis. In this study, we adopted three machine learning approaches to explore nonlinear relationships among experimental variables and low salinity effect. The best machine learning models achieved "very strong" relationships with significantly reduced MSE, particularly, random forest (RF) models perform best in four of five groups of variables, and support vector machine (SVM) fits the remaining one best. These models could be used to predict the performance of LSWF in sandstone porous media.