Deep Learning-Based Prediction of Hydrogen Dynamics and Mixing Phenomenon in Fractured Aquifers for Underground Hydrogen Storage

Abstract

Underground Hydrogen Storage (UHS) in aquifers is a promising solution. Some aquifers contain natural fractures that enhance permeability, improving injection and recovery. However, these fractures may also intensify mixing and channeling, reducing overall storage efficiency and hydrogen purity. To address these challenges, designing suitable UHS scenarios is essential to minimize hydrogen mixing and uneven distribution within the aquifer. Numerical simulations help optimize UHS operations, yet their high computational cost necessitates efficient alternatives. This study develops a grid-based proxy model using U-Net and Modified U-Net architectures to predict mixing maps and fluid flow dynamics without solving complex physical equations. The model achieves over 96% accuracy in capturing key flow behaviors like channeling and overriding while significantly reducing computational time. Results demonstrate that the optimized Modified U-Net reduces training time while maintaining prediction accuracy. The proposed framework enables rapid evaluation of different scenarios, enhancing decision-making for UHS optimization. It is applicable across various aquifer conditions, including different heterogeneities and operational settings, making it a cost-effective alternative to conventional numerical simulations.