Neural model to estimate permeability from well logs and core data

Abstract

A case study testing the effectiveness of neural networks for permeability determination in heterogeneous media using basic rock properties is presented. The dataset used consists of 213 core samples from the Morrow and Viola formations in Kansas, United States. The characterizing parameters of the cores are porosity (ϕ), water and oil saturations (Sw and So), and grain density (GD), and the additional variables from well logs are induction resistivity (ILD), gamma ray (GR) and neutron-porosity (NPHI). The neural predictions are compared with permeability values obtained from three semi-empirical models (Timur, Coates, and Pape) widely used in reservoir characterization. It is concluded that the neural network provides the best overall prediction quantified by the highest correlation coefficients (R and R2) far above those achieved with conventional methods in formations with rock heterogeneity and complex diagenetic nature. Applying Timur’s method R was 0.58 and R2 was 0.343, for Coates’ model R was 0.60 and R2 0.365 and for Pape’s model R was 0.60 and R2 was 0.372, while for the neural model, 0.97 and 0.94 were obtained for R and R2, respectively.