Machine learning-based prediction of physical parameters in heterogeneous carbonate reservoirs using well log data

This study explores the application of machine learning techniques for predicting permeability, porosity, and flow zone indicator (FZI) in carbonate reservoirs using well log data, aiming to overcome the limitations of traditional empirical methods. Six machine learning algorithms are utilized: support vector machine (SVM), backpropagation (BP) neural network, gaussian process regression (GPR), extreme gradient boosting (XGBoost), K-nearest neighbor (KNN), and random forest (RF). The methodology involves classifying pore-permeability types based on the flow index, leveraging logging curves and geological data. Models are trained using seven logging parameters—spectral gamma rays (SGR), uranium-free gamma rays (CGR), photoelectric absorption cross-section index (PE), lithologic density (RHOB), acoustic transit time (DT), neutron porosity (NPHI), and formation true resistivity (RT)—along with corresponding physical property labels. Machine learning models are trained and evaluated to predict carbonate rock properties. The results demonstrate that GPR achieves the highest accuracy in porosity prediction, with a coefficient of determination (R²) value of 0.7342, while RF proves to be the most accurate for permeability prediction. Despite these improvements, accurately predicting low-permeability zones in heterogeneous carbonate rocks remains a significant challenge. Application of cross-validation techniques optimized the performance of GPR, resulting in an accuracy index (ACI) value of 0.9699 for porosity prediction. This study provides a novel framework that leverages machine learning techniques to improve the characterization of carbonate reservoirs.