University of Wales College of Cardiff, United Kingdom

Neural networks offer an attractive way of using downhole logging data to infer the lithologies of those sections of ODP holes from which there is no core recovery. This is best done within a computer program that enables the user to explore the dimensionality of the log data, design the structure for the neural network appropriate to the particular problem and select and prepare the log-and core-derived data for training, testing and using the neural network as a lithological classifier. Data quality control and the ability to modify lithological classification schemes to particular circumstances are particularly important. We illustrate these issues with reference to a 250 m section of ODP Hole792E drilled through a sequence of island arc turbidites of early Oligocene age. Applying a threshold of >90% recovery per 9.7 m core section, we have available about 50% of the cored interval that is sufficiently well depth-matched for use as training data for the neural network classifier. The most useful logs available are from resistivity, natural gamma, sonic and geochemistry tools, a total of 15. In general, the more logs available to the neural network the better its performance, but the optimum number of nodes on a single "hidden" layer in the network has to be determined by experimentation. A classification scheme, with 3 classes (claystone, sandstone and conglomerate) derived from shipboard observation of core, gives a success rate of about 76% when tested with independent data. This improves to about 90% when the conglomerate class is split into two, based on the relative abundance of claystone versus volcanic clasts.