Committee Members Information

• Ph.D. Advisor: Prof. Bicheng Yan
• External Examiner: Prof. Mustafa Onur
• Committee Chair: Prof. Hussein Hoteit
• 4th Committee Member: Prof. Omar Knio

Abstract

Reservoir simulation is an invaluable tool for optimizing geo-energy recovery and geo-storage systems. Despite advances in numerical methods and high-performance computing (HPC), realistic reservoir models remain computationally prohibitive for workflows such as assisted history matching, uncertainty quantification, and optimization, which require numerous forward simulations. This motivates the development of surrogate models that approximate fine-scale models with reduced computational cost. Hybrid physics–data driven reduced order modeling methods have attracted growing interest because they combine physical interpretability with computational efficiency. Among them, the coarse-grid network (CGNet) accelerates simulation by representing the reservoir on a coarsened, nonuniform grid that reduces the number of cells while retaining essential flow paths. Flow equations are still solved with finite-volume discretization on this reduced graph, ensuring that fundamental physics is preserved. Compared with purely data-driven models, CGNet achieves faster evaluations while preserving physical consistency and can be efficiently calibrated against well data or reference simulations. The contributions of this work extend CGNet to multiple challenging applications. First, CGNet is applied to geological CO₂ sequestration with brine extraction, demonstrating accurate plume prediction and efficient well-control optimization for pressure management. Second, fractured reservoirs are addressed through two extensions: explicit diagonal connections for sparse fractures (CGNet-Frac) and a dual-continuum CGNet (DC-CGNet) for densely fractured systems, calibrated via ensemble data assimilation. Third, a super-resolution deep neural network (SR-DNN) is developed to enhance coarse-grid simulations by mapping coarse pressure and saturation fields to fine-scale resolution, improving predictive accuracy for multiphase flows. Finally, a deep learning inference network is coupled with differentiable surrogates to enable deterministic history matching, assimilating multi-source production and monitoring data to efficiently estimate heterogeneous model parameters. Overall, these developments establish a versatile hybrid surrogate modeling framework that significantly accelerates reservoir simulation while maintaining physical fidelity which enables rapid optimization, history matching, and performance forecasting across diverse subsurface processes, including geological CO₂ storage and flow in fractured reservoir.