On the application of neural network regression for density log construction: comparisons with traditional empirical models

Abstract

Determination of the reflection coefficients is a key element to a well-to-seismic tie, and the density log has great petrophysical importance as it is used to calculate the acoustic impedance and the reflectivity log. Many authors have developed empirical relations to determine the bulk density log from other logs information such as compressional velocity and shale volume fraction. However, as machine learning (ML) techniques have advanced, many works have used them to solve problems involving regression and classification in well-log data. The primary goal of this study is to develop Artificial Neural Network (ANN) regression models which predict the density log and use other logs as input, and then compare them to existing empirical models and find out which one provides the best fit. Two ANN models were developed, and statistical analysis was used to compare both to empirical models, such as calculating the mean squared error, relative error, and correlation factor. When compared to empirical models, both ANN models had smaller errors and higher precision on the fit.