Predicting permeability in sandstone reservoirs from mercury injection capillary pressure data using advanced machine learning algorithms

Abstract

Determining the permeability of the reservoir in the absence of well logs and core analysis data is a challenge in the oil and gas industry. Even though correlations such as Winland and Pittman exist, they often fail to provide an accurate permeability value. This study utilized the mercury injection capillary pressure data from published literature to determine the permeability in sandstone reservoirs. The dataset included parameters such as pore throat radius at various mercury saturations (25%, 35%, 50%, and 75%), along with permeability and porosity determined through laboratory experiments. Different machine learning techniques, namely, LASSO regression, ridge regression, support vector regression (SVR), random forest (RF) regression, decision tree (DT) regression, K-nearest neighbor (KNN) regression, gradient boosting, Ada Boost, and multilayered perceptron (MLP) were used to determine permeability values form porosity, pore throat radii, and pore throat sorting data. Sixty-three samples were randomly divided into training and test sets, out of which 75% were used for training both the models while 25% were used to test them. The regression coefficients suggested that pore throat radius at 75% saturation (r75) had the highest influence on the permeability values, followed by porosity (phi) and r50. It was noted that as the learning rate increased, the root mean squared error (RMSE) gradually reduced from 48.9208 to 47.2889 for ridge and LASSO-normal, while for ridge and LASSO-polynomial 99.97 to 52.2629. Various models and correlations have been developed in previous studies; however, the lithological characteristics of reservoir rock vary with location and subsurface factors. The novelty of this study lies in its integration of machine learning models with mercury injection capillary data for accurate permeability predictions, addressing the limitations of traditional correlations and offering a reliable method for characterizing sandstone reservoirs in the absence of well log data and evaluating the flow behavior of reservoir fluids within the porous media.