A Hybrid Ensemble Learning Framework for Permeability Prediction Using Integrated Petrophysical Logs, Core Measurements, and Drill Stem Test Data

Permeability is one of the most fundamental petrophysical parameters governing reservoir quality assessment, fluid-flow simulation, and hydrocarbon field development planning. Although conventional determination methods, such as core analysis and well testing, offer high measurement accuracy, their operational constraints, elevated costs, and limited vertical coverage have rendered data-driven intelligent approaches indispensable in reservoir engineering practice. In this work, a comprehensive framework built upon machine learning and hybrid Ensemble Learning techniques is presented for permeability estimation, integrating petrophysical well-log data, core measurements, and Drill Stem Test (DST) records acquired from a producing well in an Iranian gas field. To examine reservoir behavior from two complementary standpoints, the available data were organized into two principal subsets: a DST–Log dataset and a Core–Log dataset. Permeability estimation within the DST–Log subset was formulated as a classification problem, whereas the Core–Log subset was treated as a regression task. Following a preprocessing sequence comprising outlier removal, normalization, depth synchronization, and feature selection, several machine-learning algorithms—namely Logistic Regression, K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Decision Tree, and hybrid Voting/Stacking ensembles—were developed and evaluated. In addition, SHapley Additive exPlanations (SHAP) analysis was employed to enhance model interpretability and quantify the contribution of individual input features. The results indicate that tree-based algorithms and hybrid ensemble models consistently outperformed standalone learners across both classification and regression tasks. In the classification stage, the SVM, Decision Tree, and Ensemble models achieved accuracies in the range of 0.99–1.00, whereas Logistic Regression exhibited comparatively weaker performance in zones characterized by lithological overlap. In the regression stage, the Ensemble and Decision Tree models yielded the lowest prediction errors and the greatest stability in reconstructing continuous permeability profiles. Comparative analysis further demonstrated that hybrid Ensemble Learning, by integrating linear, nonlinear, local, and hierarchical learning mechanisms, substantially improved generalization capability, reduced prediction uncertainty, and enhanced robustness against noisy, heterogeneous petrophysical data. Feature-importance analysis revealed that resistivity logs (the RLA series, RXOZ, and RT_HRLT), gamma-ray indicators (CGR and SGR), and porosity–density parameters (NPHI and RHOB/RHOZ) exert the dominant control on permeability prediction. Collectively, these findings demonstrate that coupling dynamic well-testing data with static petrophysical measurements, within an integrated machine-learning and Ensemble Learning framework, provides a reliable and interpretable basis for permeability estimation in structurally complex, heterogeneous reservoirs. The proposed methodology holds practical relevance for reservoir characterization, well-completion design, production management, and future field-development planning.