ENHANCED MACHINE LEARNING APPROACHES FOR PREDICTING COMPRESSIONAL SONIC LOGS IN THE OGLET EN NASSER FIELD, ALGERIA

Abstract

Compressional sonic logs are indispensable for subsurface characterization, reservoir evaluation, and wellbore stability analysis, yet their acquisition is frequently constrained by operational limitations, high costs, and logistical challenges. This study addresses this gap by developing a machine learning framework to predict compressional sonic logs using conventional well log data from five wells. The methodology encompasses rigorous data preprocessing, including outlier removal, missing data imputation, and feature engineering to enhance predictive signal quality. Feature selection is performed to identify the most influential input variables, followed by training and validation of multiple regression algorithms: Random Forest, CatBoost, XGBoost, K-Nearest Neighbors, Support Vector Machines, and Deep Neural Networks. Hyperparameter tuning is systematically applied to optimize each model's performance, with evaluation conducted using correlation coefficients and root mean square error metrics. The ensemble-based models demonstrate the strongest predictive capability, achieving correlation coefficients between 83.6 and 88.4 percent and RMSE values ranging from 3.645 to 3.578. Additionally, the study confirms that proper input scaling is essential for distance-based and neural network models, while blind well testing further validates the generalizability and reliability of the predictions. This workflow offers a robust, cost-effective alternative for sonic log estimation in reservoir management and geomechanical studies