Compressional and Shear Sonic Log Determination Using Data-Driven Machine Learning Techniques

Shear (Vs) and compressional (Vp) sonic logs are important parameters required in reservoir exploration, development, recovery, geothermal operations, and fluid sequestration. Before the invention of the Dipole Sonic Log, older wells did not measure these essential parameters. Due to cost, these logs are not ubiquitous in all wells, especially Vs logs. Many researchers developed empirical correlations to predict Vs from Vp but have some limitations based on mineralogy, fluid types, porosity, and others. Due to the availability of massive data and technological advancements, machine learning methods have proven successful in measuring subsurface parameters from wireline logs. Although wireline logs carry vital information, domain knowledge is required to select relevant logs for model training since model performance depends on data. This study combined expert analysis with literature to select relevant logs for predictions. A multi-output Bayesian Optimised Extra-Trees (BO–ET) model is proposed to predict Vs and Vp. This research used two well data for training and testing and a third for deployment. ET surpassed the Decision Tree and Random Forest models with error metrics of 1.6 μs/ft MAE, 0.01 MAPE, and 2.7 μs/ft RMSE for both Vs and Vp. The ET model was hyperparameter-tuned using Bayesian optimisation, which improved the model performance marginally. The proposed approach outperformed empirical correlations used in predicting the Vs on the test data. The model was successfully deployed on the third well, where Vp and Vs were predicted in three different formations. The proposed approach is unique to other machine learning methodologies since it was a multi-output model prediction where Vp was not used as an input for predicting Vs.