Machine learning approach in predicting water saturation using well data at “TM” Niger Delta

Accurate estimation of water saturation (Sw) is critical for hydrocarbon exploration reservoir management, as it reveals the proportion of pore spaces filled with hydrocarbons and water. Determining Sw could be challenging in the absence of core data or resistivity logs. This informed the use of machine learning (ML) techniques to predict Sw in the "TM" Field, Niger Delta, where missing resistivity log data poses a challenge. Five ML models (XGBoost, AdaBoost, CatBoost, LightGBM, and Gradient Boost) were deployed using well log data (caliper, gamma-ray, neutron, porosity, density, and shale volume) to estimate Sw at “TM” Field, Niger-Delta. The dataset includes 61,253 observations, which were split into training (70%) and testing (30%) sets. After preprocessing and correcting inconsistencies in the data, the five ML models were trained and hyperparameters tuned to optimize performance. The models were evaluated using standard statistical metrics: Root Mean Square Error (RMSE), Mean Squared Error (MSE), Mean Absolute Error (MAE), and R-squared (R²). To validate the performance of these models, predicted Sw values were compared with those estimated from resistivity logs. Likewise the predicted Sw from the five ML models were plotted against the Sw estimated from the resistivity data not used in the prediction process to validate and determine the quality of the predicted Sw from the ML models. Among the five ML models tested, XGBoost exhibited the best performance, with the highest R² value of 0.9992 and the lowest RMSE of 0.0071. Other models, such as CatBoost, LightGBM, and Gradient Boost, showed strong performance with correlation coefficients of 0.9785, 0.9732, and 0.9299, respectively, but were less accurate than XGBoost. AdaBoost, on the other hand, demonstrated the poorest performance with a correlation coefficient of 0.4381 and the highest RMSE of 0.2082. The cross plot of the predicted Sw from XGBoost's model and actual Sw from Archie's equation had the highest correlation coefficient of 0.9, providing quality prediction thereby aligning with the statistical metrics. Hence, this study has identified Xgboost to be a promising ML tool that could be used to efficiently predict Sw without the use of resistivity data at “TM” Field Niger Delta and this could be applied in other similar geological settings.