Abstract: Artificial Neural Networks as a Tool for Porosity and Permeability Prediction

Curtis A. Link, Alden H. Wright

A complex relationship exists between lithologic parameters and their spatial distribution. Conventional reflection seismology provides a means of obtaining information about structural and stratigraphic variations; however, it is the spatial distribution of properties such as porosity and permeability that is of most interest. Artificial neural networks are proving to be a useful tool for examining this relationship.

Artificial neural networks (ANNs) are computer based models capable of recognizing complex relationships among a variety of data types without having to specify the form of the relationship. The problem of predicting the spatial distribution of porosity and permeability from seismic reflection data is approached using both supervised and unsupervised ANNs.

An unsupervised neural network called a self organizing map is used to classify the target zone of a 3-D seismic survey into distinct regions based on seismic attributes such as amplitude, phase, and frequency. Supplementary training examples are chosen for input into a supervised neural network based on the unsupervised classification. These supplementary training examples provide a wider range of data types than would be possible using only training examples from available wells.

The supervised neural network, a feed-forward network, then uses seismic attributes extracted from the data set to predict porosity and permeability values at each seismic trace location after training on porosity and permeability values from core measurements and from the supplementary training examples.

AAPG Search and Discovery Article #90952©1996 AAPG Rocky Mountain Section Meeting, Billings, Montana