Machine Learning Prediction of Dead-end Pores from VTK Based Velocity Fields in Porous Media

Dead-end pores are important microstructural features in porous media because they strongly influence flow efficiency, transport behaviour, and solute retention. However, their identification remains challenging due to geometric complexity and the limitations of conventional post-processing methods. In this study, we developed a machine-learning framework for the automated prediction of dead-end pores from pore-scale velocity-field data obtained through numerical fluid-flow simulations. Two-dimensional and three-dimensional porous geometries were constructed using CAD tools and simulated under Stokes and Navier-Stokes flow regimes using the CSMP platform. The resulting pressure and velocity fields were exported in Visualization Toolkit (VTK) format, and spatial as well as hydrodynamic features were extracted for analysis. A supervised classification approach based on a Random Forest model was trained on 222,810 pore-scale data points, using spatial coordinates and pressure as predictor variables. Dead-end pores were defined as stagnant regions weakly connected to the main flow pathways and labelled accordingly. Model performance was evaluated using accuracy, precision, recall, and ROC-AUC, along with three-dimensional visualizations for physical validation. The proposed model achieved an accuracy of 98.20%, a recall of 96.19% for dead-end pore identification, and a ROC-AUC score of 0.9985, indicating excellent predictive performance and robustness. These results show that machine learning can reliably distinguish active flow channels from stagnant dead-end regions directly from velocity-field data without explicit connectivity analysis. The approach offers a scalable and objective tool for pore-scale characterization with potential applications in enhanced oil recovery, contaminant transport, and digital rock physics.