Learning Hydrogen Flow Behavior in Heterogeneous Saline Aquifers under Cyclic Operation with Physics-Constrained CNN-RNN Framework

Geological heterogeneity in subsurface reservoirs, such as spatial variability in permeability and porosity, strongly controls hydrogen plume migration and reservoir pressure evolution during underground hydrogen storage (UHS) operated under cyclic injection and withdrawal. These heterogeneities introduce significant uncertainty in system response, complicating predictability, risk assessment, and site design. In our study, with proper distributed statistical sampling of heterogeneous permeability and porosity map on a synthetic two-dimensional saline aquifer benchmark, two-phase flow numerical simulation results reveal that cyclic hydrogen recovery performance is primarily controlled by mean reservoir permeability rather than porosity, with high-permeability formations consistently achieving the highest recovery factors regardless of porosity, while mean porosity plays a secondary, weakly controlling role.

Additional step was taking for the cyclic performance evaluation under geological uncertainties. A hybrid deep-learning surrogate framework that combines convolutional and recurrent neural network components to efficiently forecast cyclic UHS behavior under geological uncertainty. Spatial heterogeneity is captured using a U-Net-type convolutional architecture, which concisely encodes and decodes static reservoir features while preserving multiscale spatial structure. Temporal dynamics are modeled using a recurrent neural network framework adapted from ConvLSTM network (Zhao et al., 2024), enabling accurate learning of pressure and gas saturation evolution across successive injection–withdrawal cycles. This recurrent structure effectively captures cycle-dependent memory effects and dynamic transitions between operational phases. To enforce physical consistency, mass-conservation constraints are embedded directly into the training loss, preventing physically implausible predictions and improving generalization.

The developed surrogate model accurately reproduces hydrogen plume migration and reservoir pressure fluctuations observed in high-fidelity simulations. Reliable interpolation within the training cycles and extrapolation to future, unseen cycles, was validated by demostrating the performance on the synthetic aquifer benchmark. The result shows the physics-constrained model consistently outperforms a purely data-driven counterpart in predicting cyclic pressure and saturation dynamics. This approach enables the upscaling of multiphysics simulation insights into computationally efficient forecasting tools, supporting near-real-time scenario evaluation and decision-making for large-scale underground hydrogen storage under uncertainty.