USING NEURAL NETWORKS TO PREDICT THERMAL CONDUCTIVITY FROM GEOPHYSICAL WELL LOGS

Abstract

We present a new approach, based on neural networks, to predict the thermal conductivity of sedimentary rocks from a set of geophysical well logs. This method is calibrated on Ocean Drilling Program (ODP) data, which provide several thousands of conductivity measurements combined with five geophysical well logs (sonic, density, neutron porosity, resistivity and gamma ray). This data set is used to train multilayer perceptrons (MLP) and to find an empirical relationship between well logs (MLP inputs) and thermal conductivity (MLP output). Validation tests suggest that MLP provide better estimates of thermal conductivity (within ~15 per cent confidence level) than classical linear models, and still give satisfying results with sets of only four well logs if neutron porosity is included. In two ODP sites (863B and 1109D), MLP's predictions are compared to conventional 'mixing' methods. Although this latter technique gives reliable results provided that rocks description is precise enough, the MLP is more straightforward, does not need any extra parameter and makes predictions in good agreement with the experimental trends. This method will be useful in the estimation of heat flow from data acquired in scientific and industrial boreholes. © 2006 The Authors Journal compilation © 2006 RAS.