Summary Accurate prediction of permeability in carbonate reservoirs remains a significant challenge due to strong heterogeneity, diagenetic overprinting, and the limitations of empirical log-based models. In this study, a physics-guided machine learning (PGML) framework is developed that integrates empirical petrophysical knowledge into data-driven models to improve permeability estimation. The approach leverages the discrepancy between core-measured and nuclear magnetic resonance (NMR)-log-derived permeability as a domain-specific constraint. Machine learning (ML) models are trained to learn this discrepancy from conventional well-log inputs and then correct NMR permeability in a nonlinear, facies-sensitive manner. Unlike conventional workflows that apply uniform linear corrections, the proposed method explicitly accounts for depth-dependent and lithofacies-driven variability. The framework is applied to 365 core/NMR pairs from the Mishrif Formation in the Majnoon Field, southern Iraq, and evaluated using three ensemble algorithms: extreme gradient boosting (XGBoost), categorical boosting (CatBoost), and random forest (RF). Rigorous k-fold cross-validation and hyperparameter tuning ensure robust model assessment. Results show that RF achieves the highest performance R2 = 0.924, root mean square error (RMSE) reduction of 44.79% compared with both baseline ML and raw NMR permeability. The analysis further reveals that discrepancies are most pronounced in tight diagenetically altered intervals, highlighting the physical rationale for PGML corrections. Unlike previous methods that applied uniform linear corrections to NMR permeability, our approach introduces facies-sensitive, nonlinear corrections learned from well logs. This results in significant error reduction and more reliable permeability estimates across heterogeneous carbonate units, demonstrating the potential of PGML for practical reservoir characterization and improved decision making.
Khassaf et al. (2026) studied this question.