Neural Modelling of Core Permeability Data

In this paper we show a neural network based method to develop a relationship between core permeability data and a set of other well logs to complete permeability information in well areas that do not contain that information. As an alternative to the common multi layer perceptron we use a special network architecture that has several advantages in approaching the optimal network size. In combination with parallel learning, automatic network growing and sophisticated stopping criteria that architecture allows an efficient and robust estimation of the optimal network for a particular problem. We use resistivity, gamma-ray, density, neutron porosity and the p-wave sonic logs as model input yielding a heuristic permeability model that allows completion of information in the area of interest yielding a synthetic permeability log. In addition convolutional network input in combination with principal component transform enhances the quality of the resulting model.