Application of Supervised Machine Learning Technique to Investigate the Effect of Brine Hardness on Polymer Bulk Rheology

Abstract

Abstract Polymer flooding is one of the most successful methods of EOR and polymers are extensively used in EOR applications. A large number of factors affect the viscosity and rheological properties of polymers. These include but not limited to the effects of salinity, temperature and shear rate on the viscosity of polymer solutions and an exhaustive evaluation is generally required to evaluate the effects of these parameters on polymer rheology under various conditions. However, extensive manpower is required to run such experiments could hinder their applicability. Previously, several efforts have been made to develop a mathematical relationship for polymer rheology, which are limited only to a few parameters. This have prompted a pressing need for new physcis based data driven concepts and tools that can embrace the patterns embodied in these rich polymer bulk rheology data sets. In this study, an extreme gradient boosting based machine learning model is developed to predict responses to changes in monovalent and divalent ions on polymer rheology, which will provide an alternative approach. In this works, an extensive series of polymer rheology experiments are performed and the effect of concentration, temperature, shear rate and salinity on polymer viscosity are observed. The experimental data has been divided randomly in two different sets. The first set of experimental data has been used to develop ascalable end-to-end tree boosting system based machine learning model. The efficiency of the machine learning model in predicting the polymer rheology has been validated with the second set of experimental data. The machine learning model is later used to investigate the effect of individual ions temperature, brine hardness on polymer rheology. The accuracy of the machine learning model is measured using root-mean-square error (RMSE), relative error (RE), and coefficient of determination (R2). A blind test of comparing predicted results with second set of experimental data have shown that the developed predictive model is capable of predicting polymer rheology with a maximum error of 2 %, which is within the range of experimental artifact. Based on the present work it has been observed that brine hardness, with, divalent cations such as, Ca+2 and Mg+2 can significantly affect the polymer rheology. This observation can be attributed to the charge shielding effects. The HPAM based polymers and other anionic polymer have negative charges on their backbone. The long chains of polymers stretch straight in solutions due to mutual repulsion of charges, and this straight formation leads to the viscosity of polymers in brine. However, in the presence of divalent cations, these chains become curled up in ball shape formation due to the distribution of positive charges and hence, polymer losses its viscosity in solutions. This phenomenon is termed as charge shielding effect. In addition to this, the machine learning model is used to study brine hardness at varying temperature, concentration, salinity and shear rate. Finally, the machine learning model is used to predict the required polymer concentration for a given high temperature and high salinity reservoirs. Application of the proposed model will provide a clear pathway to screen polymer among wide range of reservoir conditions in order to find most suitable polymer to maximize oil recovery.