Optimizing support vector machines for enhanced permeability prediction in sandstone reservoirs using systematic and heuristic hyperparameter tuning

The accurate prediction of permeability in oil and gas reservoirs is crucial, particularly in sections where direct measurements are unavailable. This study aims to enhance permeability estimation in sandstone reservoirs by optimizing Support Vector Machine (SVM) models through two hyperparameter tuning approaches: systematic search (SS-SVM-HPO) and heuristic search (HS-SVM-HPO). The models were trained on borehole data, incorporating input variables such as Gamma Ray (GR), Neutron Porosity (Nφ), Density (DEN), Resistivity (RES), Total Porosity (PHIT), and Water Saturation (SW). Cross-validation was employed to ensure model robustness. Lithofacies were classified first, followed by permeability estimation using metrics like Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). The HS-SVM-HPO model outperformed the others, achieving an MAE of 6.85 mD and RMSE of 8.32 mD in regression tasks. When applied to an uncored well, it exhibited a strong correlation with permeability derived from porosity assessments, suggesting its effectiveness in predicting reservoir qualities. This study presents novel insights into hyperparameter optimization techniques for SVM in reservoir characterization, offering a reliable approach to permeability estimation in uncored sections, where core data is unavailable.