Physics-Informed Neural Networks For Accurate Reservoir Simulation And Enhanced Oil Recovery Prediction In Heterogeneous Formations

Reservoir simulation is fundamental to optimizing hydrocarbon recovery, yet traditional numerical methods face significant challenges in computational efficiency, handling sparse data, and modeling complex multiphase flow in heterogeneous reservoirs. This study addresses the research gap in developing robust, physics-constrained surrogate models that maintain physical consistency while reducing computational costs. We propose a hybrid Physics-Informed Neural Network (PINN) framework, specifically Hard-Soft Constraint-based PINN (HS-PINN) combined with domain decomposition (PINN-DD), for two-phase flow simulation and Enhanced Oil Recovery (EOR) forecasting. The model integrates governing partial differential equations (PDEs) of porous media flow directly into the neural network loss function. It was validated on synthetic heterogeneous reservoirs and benchmarked against commercial simulators (e.g., COMSOL) and data from the Volve field. Results show exceptional accuracy with relative errors below 1% for pressure and saturation fields, while achieving up to 92% reduction in computational time compared to traditional finite difference methods. For polymer-assisted hot water flooding EOR scenarios, the model predicted incremental oil recovery improvements of 12-19%. Feature sensitivity analysis via SHAP values identified permeability heterogeneity and injection rates as dominant factors. This work provides a scalable, interpretable AI-driven approach for real-time reservoir management and uncertainty quantification, bridging data-driven and physics-based modeling. Future extensions include integration with digital twins for full-field applications.