Physics-Informed AI for Real / Near Real-Time Reservoir Saturation Prediction & Adaptive Field Optimization

Abstract

The invention relates to a hybrid methodology for predicting water saturation (Sw) in subsurface reservoirs by combining physics-based fractional flow models with data-driven learning. Well-level water-cut is forecast using LSTM recurrent networks and transformed to saturation via inverse Buckley–Leverett or Koval formulations. An invertible neural network generates residual corrections subject to physical boundary conditions and is calibrated using high-weight saturation measurements from RST / OH logs. Spatial saturation maps are produced via radial basis function interpolation, enabling near real-time field-wide updates, uncertainty quantification, and adaptive optimization. The workflow supports parameter estimation (e.g., Koval factor, Corey exponents) by differential-evolution / Trust-Region methods and provides four variants to accommodate data-rich and data-sparse settings. (Figure: Fig. 1)