Machine learning-based hydrocarbon saturation prediction from well logs, ‘IK’ offshore field, Niger Delta Basin

This study investigates the application of machine learning (ML) algorithms for predicting hydrocarbon saturation using conventional well-log data in a Niger Delta offshore field. A preprocessed dataset of 10,825 samples from five wells was used. Four wells (IK-1, IK-2, IK-3 & IK-5) comprising 10,055 samples were split into 80% (8,044 samples) for training and 20% (2,011 samples) for validation, while IK-4 (770 samples) was withheld as an independent blind well to assess model generalization. Linear Regression (LR) and Random Forest (RF) models were developed in Google Colab to predict hydrocarbon saturation from resistivity, neutron porosity, and bulk density logs. The LR model achieved a low R2 of 0.3604, RMSE of 0.0856, and MAE of 0.0627, confirming its inability to capture the complex, non-linear relationships within the reservoir data. In contrast, the RF model demonstrated exceptional predictive performance, achieving an R2 of 0.9947, RMSE of 0.0078, and MAE of 0.0014 on the validation set, and maintained robust generalization on the blind well with an R2 of 0.9975. These findings confirm that ensemble-based ML algorithms, particularly RF, provide accurate, scalable, and transferrable solutions for hydrocarbon saturation prediction in heterogenous offshore fields.