Since the early stage of the oil industry, many authors have recognized the importance of capillary pressure measurements using the mercury injection method (MICP) method in estimating permeability values. For that reason, numerous permeability models have been developed and proposed in literature. MICP is a very popular technique to determine pore size distributions and pore-throat properties. The capillary pressure profiles determined using this method are influenced generally by certain parameters that are controlled by permeability, including: degree of sorting, pore size distributions and pore throat properties.
Since MICP offers direct relation to permeability-related information, a good estimation of permeability should be obtained. Most of the available correlations produce high errors when compared to actual permeability measurements. To address this issue, a feedforward neural network (ANN) model was developed to predict permeability from the MICP measurements. The neural network consists of two hidden layers with 15 neurons each and one output layer.
A dataset of 206 core samples were used to train the ANN model. The dataset was divided into three sets: 70% for training, 15% for internal validation, and 15% for blind testing. A variety of parameters — porosity, displacement pressure, Swanson parameter, Winland parameter, Dastidar parameter, Pittman parameter, and Purcell integration — were extracted from the MICP data and inputted to the ANN model. In the analysis to evaluate the developed model against conventional models, graphical and statistical comparisons were used to determine the best technique for use.
Comparison between permeability models available in literature and the developed ANN model indicated significant improvement with the proposed model. Error measures — including: maximum absolute percent error (MAE), average relative percent error (ARE), average absolute relative percent error (AARE), correlation coefficient (R2), standard deviation (SD) and root mean squares (RMSE) — were used as the basis for interpreting the results and evaluating the comparative performance of the models. The developed ANN model demonstrated a reduction of about 50% in AARE and RMSE. Correlation coefficient also improved from 0.91 for the best conventional model to 0.98.