Deep learning-based surrogate modeling for underground hydrogen storage

Underground hydrogen storage (UHS) is critical for integrating intermittent renewable energy sources by storing excess energy as hydrogen during surplus periods and retrieving it during shortages. Effective UHS design requires accurate prediction of hydrogen plume migration and reservoir pressure evolution, typically achieved by high-fidelity numerical simulations. Although these physics-based simulations are accurate, they are computationally expensive and unsuitable for rapid decision-making. To address this, we develop efficient surrogate models for UHS prediction using Swin-Unet, a transformer-based deep learning architecture with a U-Net structure. Our results show that Swin-Unet accurately predicts the spatiotemporal evolution of hydrogen saturation and reservoir pressure in heterogeneous depleted gas reservoirs, offering a fast and reliable alternative to traditional simulations. Compared to surrogate models based on U-Net and Segmentation Transformer (SETR), Swin-Unet achieves higher pressure prediction accuracy while maintaining similar training costs. For hydrogen saturation prediction, Swin-Unet achieves higher accuracy than SETR and comparable accuracy to U-Net, while reducing GPU memory usage by about two-thirds and training time by approximately 75% relative to U-Net. This is the first study to apply Swin-Unet to surrogate modeling of UHS, demonstrating its potential as an accurate and efficient approximation for traditional physics-based simulations. Swin-Unet enables accurate and efficient UHS prediction in heterogenous geological formations, supporting sensitivity analysis, uncertainty quantification, and operational optimization for future UHS projects.