Compressional Sonic Log Prediction Experiment Using Machine Learning: Algorithms and Feature Selection

Sonic logs are important parameters of subsurface rock properties, and are used in various stages of oil and gas exploration as well as field development. However, these measurements are sometimes missing in certain depth intervals due to tool failure. In this study, we are comparing the effects of machine learning methods and feature selection on the predictive accuracy of compressional sonic log (DTC). We utilized the data of five wells and studied the comparative performance of Artificial Neural Networks, Regression Trees, Support Vector Machines, and Random Forest on DTC prediction. Random forest had the highest correlation coefficient and lowest mean absolute percent error, and was thus used to test the effect of feature selection on prediction accuracy. We used different input features in three scenarios: the first used only wireline data, the second used only drilling data, and the third combined both. We concluded that wireline data is sufficient to predict DTC with high accuracy. Using drilling data alone would be useful if information on rock strength were needed in real-time, but should not be relied upon for accurate prediction. Combining both and increasing the number of features did not improve prediction accuracy.