Abstract

Underground hydrogen storage (UHS) can provide long-duration and large-scale energy storage. Field observations reveal that injected hydrogen can stimulate subsurface microbial reactions, resulting in compositional changes in reservoir fluids and hydrogen loss. The primary objective of this dissertation is to model and simulate the complex interactions of compositional multiphase flow, microbial reactive transport, and geochemical reactions to support safer and more efficient UHS design and operation 

To accurately capture phase and compositional behavior in UHS, a hybrid flash calculation model is developed in Chapter 2. To accelerate the flash calculations, two deep neural network (DNN) models are constructed for phase stability analysis and initialization of equilibrium ratios, respectively. The proposed model shows good agreement with experimental hydrogen solubilities and achieves an approximately twice speedup compared with methods of Wilson and compositional space.  

Building on this thermodynamic framework, a new reservoir simulator is developed in Chapter 3 by fully coupling two-phase, multi-component compositional flow with microbial reactive transport. Spatial discretization is performed using a control volume finite difference (CVFD) scheme, while temporal discretization is achieved using a fully implicit first-order backward Euler method. A modified double-Monod model is implemented to represent the growth and decay kinetics of microbial populations. The simulator is systematically verified and validated: the compositional flow module shows close agreement with a commercial simulator in both primary and secondary variables, while the microbial reaction module successfully reproduces published experimental observations.  

To further characterize microbially mediated bio-geochemical processes, a thermodynamic-kinetic batch reaction model is developed in Chapter 4. It integrates the thermodynamic feasibility of microbial growth via Gibbs free energy constraints with kinetic rate expressions, and accounts for energy allocation among microbial growth, decay and maintenance metabolism. This framework provides a more detailed thermodynamic and kinetic characterization of the underlying microbial and geochemical processes.  

From an operational and economic perspective, a surrogated model-based optimization framework is developed in Chapter 6 by integrating compositional reservoir simulation, deep neural network (DNN) surrogate modeling, and optimization algorithms. Using dataset generated from reservoir simulations, convolutional neural network–bidirectional long short-term memory–attention (CNN-BiLSTM-Attention) models are trained as surrogates, which are subsequently incorporated into deterministic and stochastic optimizations based on genetic algorithms to maximize net present value (NPV). The proposed approach achieves an approximate 4878 times speedup compared to that replying solely on physical reservoir simulation.