Design of Artificial Neural Network for predicting the reduction in permeability of porous media as a result of polymer gel injection

Abstract

Cross-linked polymer gel is widely used in the oil and gas industry to block high permeability conduits and reduce water cut. The complex nature of this fluid, especially regarding flow in porous media, makes its numerical simulation very time-consuming. This study presents an approach to designing an Artificial Neural Network (ANN) model that could predict the permeability reduction caused by injecting polymer gel into a 2D rock sample. Our methodology consists of two main parts: numerical simulation and ANN model building. Considering the advantages of the Lattice Boltzmann Method (LBM) this approach is used to model the injection of polymer gel in porous media. Using this model, more than 20,000 simulations were performed which resulted in highly unbalanced dataset, so an innovative approach for balancing regression dataset is also proposed in detail in this paper. The final constructed ANN model could predict the permeability reduction in a fraction of a second with less than 2.5% Mean Absolute Error (MAE). The result indicates the importance of balancing datasets to obtain a reliable prediction from ANN. Also, it should be mentioned that gelation parameters had the most significant impact on the value of permeability reduction, with mean absolute SHapley Additive exPlanations (SHAP) values of 20 and 12.5 for TDfactor and Threshold, respectively.