Estimation of Vertical Permeability of Hugin Sandstone from Petrophysical Well Logs Using Ensemble Methods: An Enhanced Machine Learning Approach

Understanding the directional distribution of permeability is crucial for oil and gas exploration. In this study, well log and core data were used to predict continuous vertical permeability (K_v) logs. The predictions were performed for two wells drilled through the Hugin sandstone in the Volve field. Conventional well log and corresponding vertical core permeability measurements were utilized for training machine learning (ML) models. Data enhancement techniques, such as log smoothing, data transformation, and outlier removal were used. Thereafter, ensemble models, such as Random forest (RF), gradient boosting (GB), extreme gradient boosting (XGBoost), and adaptive boosting (AdaBoost) were used with decision trees, which are the base learners. Multilinear regression models were used for comparison of K_v values from ML algorithms and core measurements. Hyperparameter tuning was performed using grid search to obtain the best learning parameters for model optimization. The trained predictive models were used to predict K_v in two wells. Four metric scores, i.e., coefficient of determination (R²), mean squared error (MSE), mean absolute error, and root mean square error were used for model evaluation. The obtained R² of RF and AdaBoost models are 0.94 and 0.92, respectively. Moreover, the R² of 0.99 for GB and XGBoost are overestimates confirmed by their high MSE values. The standard deviation of R²’s obtained from four-fold cross-validation indicates that GB and XGBoost are less stable compared to RF and AdaBoost. The RF ensemble model outperformed others. Moreover, a multilinear regression equation relating K_v to various raw logs has been provided.