Fusing Data-Driven Insights with Physics for Underground Hydrogen Storage

Abstract Underground Hydrogen Storage (UHS) in aquifer reservoirs is pivotal for stabilizing the supply of renewable energy, addressing its inherent variability. As UHS technology evolves, the need for analyses that capture the complex interactions of hydrogen within subsurface environments becomes increasingly critical. To meet this requirement, we utilize the Eclipse 300 compositional simulator with the GASWAT option to generate high-fidelity datasets, which model the intricate gas-aqueous phase equilibria essential for understanding hydrogen behavior underground. These datasets, while fundamental, are supplemented by our Physics-regularized Fourier-Integrated Hybrid Deep Neural Framework (PR-F-IHDNF) to enhance predictive capabilities. This deep learning-based surrogate model integrates convolutional LSTM, convolutional neural networks, and Fourier neural operators, all regularized with the Hydrogen-Water Mass Balance Equation, to predict the evolution of pressure and hydrogen saturation over time during injection and production cycles. Our case study of the Fenton Creek field involved detailed reservoir modeling based on a grid of 97 × 18 × 35 cells, each measuring 121×136×2.8 ft. Although the entire grid was used to generate comprehensive simulation data, we concentrated on a sector grid of 44 × 11 × 11 cells for PR-F-IHDNF training to enhance computational efficiency. This sector, strategically centered around a key well, allowed us to accurately capture dynamic hydrogen behavior. Through Latin Hypercube sampling, we explored a range of reservoir properties and operational parameters, adapting our modeling techniques to the cyclical nature of hydrogen storage and retrieval. During the data generation phase, 76 simulations were completed within 48 hours. Each simulation or realization encompassed a 24-month cycle of hydrogen injection and production, initiating with 6 months of hydrogen cushion gas injection followed by alternating three-month cycles of production and injection. This sequence resulted in three complete cycles after the initial cushioning phase. PR-F-IHDNF was trained using 26 simulation realizations and validated with 15 realizations to monitor training performance and prevent overfitting. Additionally, 35 simulation realizations were used to test the trained PR-F-IHDNF, ensuring its generalization capabilities. Results from deploying the PR-F-IHDNF showed high precision, achieving an accuracy of 99.7% for pressure and 97% for hydrogen saturation across 35 test realizations—more than the 26 used in training—to robustly verify its generalization capabilities. This outcome underscores the efficacy of incorporating the Hydrogen-Water Mass Balance Equation for regularization. The mean absolute error was recorded at 10.54 psi for pressure and 0.0018 for hydrogen saturation, indicating good predictive reliability. Although training the PR-F-IHDNF required significant computational resources, with a training duration of 36 hours and early stopping implemented at 271 epochs of the planned 300, it efficiently predicts outcomes for any simulation case in less than 0.8 seconds, showcasing its practicality for real-time applications. The PR-F-IHDNF model can predict complex underground processes, making it useful for testing different scenarios and improving storage strategies. It helps identify important factors and refine operations, supporting better decisions for managing underground hydrogen storage.