UHSNet: Deep learning-based smart proxy modeling for underground hydrogen storage

Underground hydrogen storage (UHS) in saline aquifers has emerged as a pivotal strategy for large-scale, long-duration energy storage in a hydrogen-based economy, enabling seasonal balancing of supply and demand. However, simulating the multiphase flow of hydrogen and brine in porous media is computationally intensive and highly nonlinear, particularly under variable geological and thermodynamic conditions. Traditional computational fluid dynamics (CFD) approaches, while accurate, are time-consuming and often unsuitable for real-time decision-making or large-scale scenario evaluations. To overcome these challenges, an efficient AI-based proxy model, UHSNet, has been developed for spatio-temporal hydrogen saturation estimation in UHS systems. The architecture integrates dilated convolutional layers, residual connections, and a hybrid Huber-MAE loss function to enhance predictive performance and generalization. A high-fidelity dataset was generated using a flexible CFD simulator, enriched with laboratory data and thermodynamic principles, and sampled using the Latin Hypercube method to ensure comprehensive data coverage. Four image-to-image architectures were trained and evaluated using quantitative metrics and visual analysis. UHSNet achieved a minimum 15 % reduction in mean absolute percentage error (MAPE) compared to baseline CNNs and Unet variants, along with faster training convergence. On the test dataset, the model reached 4.65 % MAPE and demonstrated prediction times under 1 s—up to 104 × faster than conventional CFD simulations. These outcomes highlight UHSNet's effectiveness as a reliable, real-time alternative to CFD for UHS applications, contributing to the optimization of underground hydrogen storage and supporting energy system planning in the renewable energy sector.