Application of the radial basis function neural network to the prediction of log properties from seismic attributes

Abstract

In this paper, the radial basis function neural network (RBFN) is used to predict reservoir log properties from seismic attributes, and is compared to the generalized regression neural network (GRNN), discussed by Hampson et al. (2001). Both of these methods are related to the probabilistic neural network (PNN), which uses a Gaussian kernel estimator based on the distance between points in seismic attribute space. Our goal was to see if there are situations in which the RBFN could improve on the results found using the GRNN. The methodology consisted of first training the neural network at each well location, to find the optimum set of seismic attributes and weighting coefficients, and then applying the trained network to a 3D seismic volume. Each neural network was applied to the Blackfoot 3D seismic volume, which was recorded over a Cretaceous channel sand in central Alberta. Our results showed that, although the training results were quite close for each method, the RBFN approach generally produced higher-frequency results, especially as the number of training values was reduced. By computing the least-squared error between the predicted samples and the known training samples, we were also able to demonstrate the improvement of the results of RBFN over GRNN as the number of training samples decreased.