Prediction of sonic log using petrophysical logs via machine learning technique

Abstract

The sonic log is the pivotal parameter for the reservoir description and fluid identification and is extensively applied in determining mechanical rock properties for rock physics, quantitative seismic interpretation, and geomechanics application. There is frequently a paucity of shear wave velocity (Vs) data in oil and gas exploration wells which is relatively due to poor borehole conditions (washout), damaged tools, offset well data, and quite expensive. This paper aims to provide a solution to predict the compressional wave (Vp) and shear wave velocities (Vs) by machine learning (ML) model using the original petrophysical data from an oil and gas sandstone reservoir in the Malay Basin and build a generalisable ML model. The ML framework is based on Cross Industry Standard Process for Data Mining (CRISP-DM) workflows and Exploratory Data Analysis (EDA) as an iterative cycle to analyse and tune the algorithms. First, appropriately address the composite data and its associated uncertainties through data pre-processing. Second, set the data splitting and evaluate the prediction model’s through several regressions. Third, run an optimization algorithm to search for the best hyperparameters for the regressor to optimize the prediction. The ML method then captured the performance measure from the Coefficient of Determination (R2) of 0.96 and 0.97 for Random Forest and Decision Tree Regression, respectively, and the lowest Root Mean Square Error (RMSE) value was recorded at 0.05, which indicates the excellent model with positive correlation. It is observed that the predicted Vp (DTC logs) and Vs (DTS logs) of the ML model produced good cross-validation to the original logs with a good performance measure of 1.0 for R2 and 0.0 for RMSE. It can be concluded, based on the performance measure of each method, indicates the robustness of DTS log prediction using a quantitative measure of accuracy in scoring the predictions. It demonstrates the ML model’s ability to generalize and predict shear logs on full field size.