The Use of Artificial Neural Networks in Reservoir Permeability Estimation From Well Logs: Focus on Different Network Training Algorithms

Abstract

This article demonstrates how to estimate permeability in different levels of a reservoir using artificial neural networks. Well logs data from one of the Iranian oil fields are used as inputs to networks. In this article, results of five common training algorithms, including Levenberg-Marquart, Bayesian regularization, gradient descent, resilient back-propagation, and one step secant, have been compared. Among all training algorithms, resilient back-propagation had the best performance by a mean squared error of 0.294. Correlation coefficient of estimated permeability values derived from the resilient back-propagation algorithm versus core permeability values have been presented. In addition, the permeability is estimated by the multiple linear regression method. Results demonstrate that artificial neural network is more efficient and trustworthy in permeability estimation than multiple linear regression method.