Abstract
Accurate estimation of water saturation (Sw) is essential for optimizing oil recovery strategies and is a key component in petrophysical analyses of hydrocarbon reservoirs. Traditional Sw estimation approaches often face limitations due to idealized assumptions, dependency on core-derived parameters, and geological heterogeneity. In this study, a comprehensive dataset consisting of 30,660 independent data points was utilized to develop machine learning (ML) models for Sw prediction. Nine well log parameters-Depth (DEPT), High-Temperature Neutron Porosity, True Resistivity, Computed Gamma Ray, Spectral Gamma Ray, Hole Caliper, Compressional Sonic Travel Time, Bulk Density, and Temperature-were used as input features to train and test five ML algorithms: Linear Regression, Support Vector Machine (SVM), Random Forest, Least Squares Boosting, and Bayesian methods. To improve model performance, a Gaussian outlier removal technique was applied to eliminate anomalous data points. The models were rigorously validated using multiple training/testing data splits and ten independent runs to ensure statistical reliability. Among the tested models, SVM achieved the highest accuracy, with R(2) values of 0.9952 (test) and 0.9962 (train) and RMSE values of 0.002 (test) and 0.001 (train). These results demonstrate that ML-particularly SVM-offers a robust and accurate alternative for Sw estimation, supporting more effective reservoir evaluation and oil recovery optimization.