Prediction of porosity and lithology in siliciclastic sedimentary rock using cascade neural assemblies

A novel approach to porosity and lithology prediction in siliciclastic sedi mentary rocks is discussed here. It is based on the use of multi-element ne ural assembly with multiple inputs and a single output. In experiments with porosity prediction, the writers used one-, two- and three-dimensional inp ut vector-parameters with coordinates (1) V-p or V-s; (2) V-p and V-s, and (3) petrophysical group index, V-p and V-s. When the neural assembly was tr ained to predict lithology, the input parameters were: (1) V-p or V-s; (2) V-p and V-s, and (3) V-p, V-s, and porosity. The writers utilized the train ing data-sets containing only 20 to 30 elements. To be able to work efficie ntly with such small training sets the writers used cascading neural assemb lies specifically designed to work with small. training data-sets. Each ele ment of the neural assembly is a neural network of a simple structure. The elements of the assembly are combined in such a way that the approximation error of the assembly obtained during training session decreases with incre asing number of elements. This allows for a simple and a well-defined metho dology of estimating of a necessary number of neural elements in the assemb ly. The cost function of each element of neural assembly was taken as a sum of two terms. The first term was an estimate of the prediction error (a st andard cost function of the predictive neural net), whereas the second term was a regularization term equal to the weighted sum of the squared norms o f the transform matrixes. The neural network-based prediction of rock param eters was tested on a variety of training and test data-sets. Best results were achieved when the training data-set included representatives of all li thologies (petrophysical groups) contained in the test data-set and when th e input parameters included independent data on the type of lithology or po rosity in addition to seismic velocities. (C) 1999 Published by Elsevier Sc ience B.V. All rights reserved.