Physics-Based and Data-Driven Polymer Rheology Model

Abstract

Summary Polymer flooding is a common enhanced oil recovery (EOR) method used to increase aqueous phase sweep efficiency by increasing viscosity. Estimating polymer viscosity for given reservoir conditions (i.e., oil viscosity, temperature, and brine composition) requires intensive laboratory work. There are existing empirical models to estimate polymer bulk rheology without prior laboratory work; however, they have many coefficients, simple brine composition, and lack physics-based regression boundaries. This study benchmarks the existing polymer empirical and machine learning (ML) models against a new data-driven model with some physics basis for common synthetic polymers. We cover a broad range of polymer concentrations, temperature, salinity, and hardness with an upper limit of 5,000 ppm, 120℃, 290,000 ppm, and 33,000 ppm, respectively. The data were preprocessed through data analytics techniques, and a model was developed with some physics basis by fitting Martin’s equation for Carreau model coefficients. Our regression boundaries obey flexible polymers’ physical and laboratory behavior. We benchmarked the bulk rheological model with existing models in the literature. We used the published models’ coefficients and then tuned their coefficients for our data set for a fair comparison. We then investigated ML as a predictive tool without compromising overfitting the data using the simplest ML model (linear regression) all the way to artificial neural network (ANN) and hybrid ML models. This is the first study that comprehensively benchmarks polymer rheology models and proposes a simple, least number of coefficients, and tunable polymer-rheology model. We provide a predictive bulk rheology model that enables the user to accurately predict polymer viscosity without laboratory measurements and for a wide range of temperatures and brine compositions. Moreover, our study includes the recently common polymer SAV-10 that was not previously studied. We present a simple water viscosity model for a broad brine salinity and temperature range. Our study shows that ML techniques might provide deceptively high accuracy for small data sets, unless due diligence is done to avoid a high-variance model.