Comparative assessment of machine learning models for polymer solution viscosity prediction in enhanced oil recovery.

Polymer flooding is among the most researched and utilized methods to improve oil recovery. Polymer solution viscosity is the primary governing parameter that aids in lowering the mobility of the injected fluid and augmenting the volumetric sweep efficiency. However, the design of a polymer solution is time-consuming and requires intensive resources because of the need to design several precursors, including makeup water salinity, temperature, polymer concentration, etc., to sustain the target solution viscosity. The objective of this study was to develop a fast and reliable approach to predict the viscosity of SAV10 polymer using the necessary input parameters, including polymer concentration, shear rate, temperature, and brine salinity. Four simple machine learning techniques, namely linear regression (LR), support vector machine (SVM), decision tree (DT), and artificial neural network (ANN), were implemented with varying critical model parameters in each technique. Owing to the non-linear relationship between polymer viscosity and input parameters, a simple linear regression model was unable to capture them. The best machine learning model among the classical models used in this study was a Wide ANN model with one fully connected layer of 100 neurons. This model yielded an acceptable viscosity prediction with a coefficient of determination (R²) of 0.998 and a root mean square error (RMSE) of 0.31 for the test data. To further generalize the predictability and reduce the model error, advanced machine learning models, i.e. Gaussian Process Regressor and an ensemble machine learning stacking regressor was employed. However, the stacking regressor dramatically lowered the root mean square error by almost 40% while the model successfully generalized the unseen viscosity data and predicted the high viscosity values with considerable confidence. We developed an SAV10 viscosity predictor from routine inputs where a stacking-based ensemble resulted in an excellent match between the actual and predicted viscosities having R2 value of 0.998 with minimum RMSE of 0.208, and MAE of 0.223, while preserving accuracy on unseen high-viscosity data.