Machine Learning-Based Prediction of Formation, Facies, Porosity, and Permeability in a Carbonate Reservoir of the "GTR" Field

Abstract

Reservoir characterization is essential for understanding rock and fluid behavior in hydrocarbon field development. In the Baturaja Formation, Sunda Basin, this process i s challenging due to heterogeneity resulting from depositional and diagenetic variations. Limited core data and the high cost of conventional analysis encourage the use of machine learning (ML). This study aims to predict formation, facies, porosity, and permeability using ML algorithms and to assess the impact of feature augmentation. The dataset includes well log and core data from 13 wells. The workflow consists of preprocessing, feature selection, feature engineering, and supervised learning using Decision Tree, Random Forest, XGBoost, and KNN. Performance is evaluated using the F1-score for classification and MAE/RMSE for regression, followed by blind testing on wells HARLEY and XSR. Random Forest achieves the best formation prediction (F1-score 0.9890; blind test 0.9975) because the well data fall within the range of the training data distribution, although accuracy decreases in XSR due to differences in data distribution. XGBoost is the most accurate for facies prediction, improving from an F1-score of 0.9648 to 0.9741 after feature augmentation. For porosity and permeability, Random Forest produces the lowest errors, although permeability remains challenging in heterogeneous carbonates. Overall, ML provides an efficient and accurate approach, with Random Forest and XGBoost performing best, and feature augmentation consistently enhancing model generalization.