A Machine Learning Approach to Shear Sonic Log Prediction

Abstract

A machine learning approach to shear sonic log prediction is demonstrated. The results of this approach were compared to that of an approach based on the Greenberg-Castagna empirical method. This approach is based on supervised machine learning and is implemented in MATLAB. While the Greenberg-Castagna method is an empirical method that attempts to predict shear velocity log from compressional velocity log for various pure and composite lithologies, this approach uses, in addition to compressional velocity log as the main predictor, several other logging measurements as predictors including gamma ray, bulk density, neutron, resistivity, porosity and water saturation logs. A dataset which includes wells with recorded shear velocity logs is used to train and validate the machine learning model. A feature selection process is performed to highlight which of the logs would be good predictors of shear velocity (VS). Various regression models are then trained, and the predicted values compared to the actual for the various models by their root-mean-square errors (RMSE), and the model with the smallest RMSE is chosen. Predictions are then carried out on another well within the dataset, which serves as the validation set. The results show improvement in the accuracy of the predictions over the linear regression model based on the Greenberg-Castagna method, as measured by the RMSE. The case study also demonstrates the potential of carrying out shear sonic log prediction in hydrocarbon-bearing intervals, which is a limitation of the Greenberg-Castagna method which only works in brine-saturated rocks. This approach would provide improved accuracy where shear sonic logs are absent and need to be predicted for geomechanics, rock physics and other applications. This is particularly important in older fields where shear sonic logs were never acquired in the older wells.