Predicting HPAM Polymer Yield Viscosity using Neural Networks

Abstract This work follows paper SPE- 216592-MS presented in October 2023 and gives an update of the machine learning model developed by SNF to predict HPAM polymer yield viscosity under various conditions. Major improvements to the model are as follows: The random forest algorithm has been replaced by a neural network to predict EOR polymer yield viscosity. The rheological dataset has been extended to include more than 95,000 lab measurements compared to 69,000 previously. The rheological dataset has been filtered using the DBSCAN clustering algorithm to detect anomalies in the polymer flow curves and 1,030 erroneous lab measurements could be removed from the training dataset. A more limited selection of input features has been used for the prediction with only 7 variables compared to 8 previously. Only the polymer ATBS content was retained and deemed sufficient to characterize the polymer chemistry due to strong correlations between acrylamide and ATBS contents. A new web interface was developed to deploy the model to the end user. These modifications helped improving the model's prediction accuracy while limiting the overfitting tendency. The mean absolute percentage error for the yield viscosity prediction dropped from 35.8 % down to 16.0%. A second objective of the machine learning model was to assist lab technicians with polymer screening. The model was hence adapted to select suitable polymers based on the target viscosity and reservoir conditions (reservoir permeability, temperature, brine salinity and hardness) while ranking those selected polymers to minimize the required polymer concentration and cost. In addition to those efforts, another app was developed to visualize the different polymer flow curves present in the rheological dataset while comparing them with fitted regression lines based on the Carreau – Yasuda equation. Parameters such as relaxation time λ or power law index n have been derived by fitting the Carreau – Yasuda model on 4,680 polymer flow curves.