Intelligent Predictor for Polymer Viscosity to Enhance Support for EOR Processes

Research into the use of polymers for enhanced oil recovery (EOR) processes has been going on for more than 6 decades and is now classified as a techno-commercially viable option. A comprehensive evaluation of the polymer's rheology is pivotal to the success of any polymer EOR process. Laboratory-based evaluation is critical to EOR success; however, it is also a time/capital consuming process. Consequently, any tool which can aid in optimizing lab tests design can bring in great value. Accordingly, in this study a novel predictive correlation for viscosity estimation of commonly used "FP 3330S" EOR polymer is presented through use of cutting-edge machine learning neural networks. Mathematical equation for polymer viscosity is developed using machine learning algorithms as a function of polymer concentration, NaCl concentration, and Ca2+ concentration. The measured input data was collected from the literature and sub-divided into training and test sets. A wide-ranging optimization was performed to select the best parameters for the neural network which includes the number of neurons, neuron layers, activation functions between multiple layers, weights, and bias. Furthermore, the Levenberg-Marquardt back-propagation algorithm was utilized to train the model. Finally, measured and estimated viscosities were compared based on error-analysis. Novel correlation is developed for the polymer that can be used in predictive mode. This established correlation can predict polymer viscosity when applied to the test dataset and outperforms other published models with average error in the range of 3-5% and coefficient of determination in excess of 0.95. Moreover, it is shown that neural networks are faster and relatively better than other machine learning algorithms explored in this study. The proposed correlation can map non-linear relationships between polymer viscosity and other rheological parameters such as molecular weight, polymer concentration, and cation concentration of polymer solution. Lastly, through machine learning validation approach, it was possible to examine feasibility of the proposed models which is not done by traditional empirical equations.