A Hybrid Physics-Informed and Data-Driven Approach for Predicting Water Volume in Oil Reservoirs with Explainable AI

Predicting water volume in oil reservoirs is a critical challenge for reservoir engineering, impacting production optimization and field development. Traditional methods like reservoir simulation and empirical modeling require extensive, often incomplete, geological and dynamic data. Recent machine learning advancements offer data-driven solutions but lack interpretability and physics integration. Hybrid approaches, combining data-driven techniques with physics-informed models, are gaining traction, yet their potential with explainable AI (XAI) for water breakthrough insights remains underexplored. This study compares a hybrid Physics-Informed Neural Network (PINN) with a baseline feedforward network and data-driven benchmarks (Random Forest, XGBoost) using historical production data (flow rates, pressures, porosity) from the SPE data repository. By embedding Darcy's Law, the Buckley–Leverett saturation model, and a global material balance constraint into the PINN loss function, we enforced physical consistency during training. Results show PINN reducing RMSE from 107.64 stb (baseline) to 96.24 stb and MAE from 45.58 stb to 40.99 stb, while XGBoost achieved the lowest errors (RMSE 84.14 stb, MAE 34.15 stb, R² 0.38). SHAP analysis highlighted "Calculated Sandface Pressure," "Line Pressure," and "Oil Volume" as key drivers, providing actionable insights for choke and gas lift management. This study demonstrates that integrating physics with machine learning enhances accuracy and physical reliability for reservoir decision-making. Future work will explore adaptive physics constraints and sequence-aware architectures for real-time forecasting in complex settings.