Physics-Informed Graph Neural Networks (PI-GNN) with Uncertainty Quantification for Optimizing Oil Recovery

Abstract Fractured carbonate reservoirs in the Middle East present challenges for Enhanced Oil Recovery (EOR) due to complex fracture networks, uneven permeability, and high water cut. Traditional modeling methods struggle with scalability, interpretability, and uncertainty. This study introduces a Physics-Informed Graph Neural Network (PI-GNN) with Uncertainty Quantification (UQ) to address these limitations. PI-GNN models reservoirs as graphs, embedding physical laws (e.g., Darcy's law) to ensure consistency, while UQ provides probabilistic outputs for improved decision-making. Results show 15–25% improved accuracy in pressure and production predictions, 20% better physical consistency, and fivefold faster simulations. UQ captures uncertainties in 90%+ of cases, enhancing reliability. This framework offers a scalable, interpretable, and efficient solution for optimizing EOR strategies, bridging physics-based modeling with machine learning for improved reservoir management.