Physics-Informed Machine Learning for Enhanced Permeability Prediction in Heterogeneous Carbonate Reservoirs

Abstract Physics-informed-machine-learning (PIML) techniques enhance permeability prediction in carbonate reservoirs. Accurate permeability estimation is crucial for reservoir characterization, fluid flow modeling, and oil/gas production optimization. However, traditional empirical models and conventional machine learning techniques often fail to capture the complex nonlinear relationships governing permeability, particularly in heterogeneous carbonate formations. To address this challenge, this research integrates physics-based constraints into machine learning models, thereby improving their predictive accuracy and robustness. Three tree-ensemble algorithms—Extreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), and Random Forest (RF)— are developed to predict permeability from well-log data inputs. The study introduces a discrepancy model that leverages the residuals between real core and nuclear magnetic resonance (NMR) permeability. The machine learning models are trained to predict these discrepancies and subsequently incorporate them to generate improved permeability predictions. A rigorous cross-validation process is applied to ensure model reliability, and statistical validation metrics, including adjusted R2 and root mean square error (RMSE) are calculated to assess prediction performance. The results demonstrate that incorporating physics-informed constraints significantly enhances the predictive performance of the three machine learning models. Random Forest achieved the highest permeability prediction performance (R2 = 0.908; RMSE = 16.73 mD) outperforming XGBoost and CatBoost. The PIML approach effectively minimized prediction errors associated with conventional data-driven methods, leading to permeability estimates that are more closely aligned with laboratory core measurements. The results confirm the power of PIML as a tool for improving permeability modeling in carbonate reservoirs. By embedding domain-specific physical constraints into machine learning models, this approach bridges gaps that exists between empirical data-driven predictions and fundamental reservoir physics. It offers a more reliable and generalizable permeability estimation framework. Future research should explore the development of the PIML techniques to optimize reservoir characterization in ways that can enhance field development strategies that lead to improved petroleum recovery.