محسن فلاحي

Accurate simulation of two-phase flow in porous media is the cornerstone of reservoir management an‎d emerging underground hydrogen storage (UHS) projects. Despite the high fidelity of classical numerical methods, their prohibitive computational costs in large-scale systems an‎d the physical inconsistency of purely data-driven machine learning models have necessitated a shift toward new computing paradigms. This research develops a Physics-Informed Convolutional Neural Network (PICNN) framework capable of solving the governing partial differential equations (PDEs) of oil-water flow without reliance on labeled training data. The core innovation lies in replacing the conventional automatic differentiation with a finite-volume-based formulation within the loss function an‎d integrating the Peaceman well model to accurately capture bottom-hole pressure (BHP) dynamics. The modelʹs performance was eva‎luated across three levels of complexity, including homogeneous an‎d heterogeneous reservoirs, an‎d a complex rectangular five-spot pattern at the field scale. Validation results against a reference simulator (CMG) demonstrate high accuracy, with a mean relative pressure error of less than 5%. Furthermore, the implementation of a transfer learning strategy resulted in a 3-fold acceleration in convergence across successive timesteps. Although challenges regarding the smearing of the saturation front in high-gradient regions were identified, the excellent agreement in sweep patterns proves that the proposed model is a robust an‎d efficient tool for digital twin development an‎d the management of energy transition projects.

شبيه‌سازي مخزن , جريان دوفازي در محيط متخلخل , يادگيري عميق , شبكه عصبي پيچشي آگاه از فيزيك

Reservoir Simulation , Two-phase Flow in Porous Media , Deep Learning , Physics-Informed Convolutional Neural Network (PICNN)

Mohsen Falahi

Dr. Majid Siavashi