Reproduce “Enabling Clean Energy Resilience with Machine Learning-Empowered Underground Hydrogen Storage” on OpenResearch

About this paper

Alvaro Carbonero, Shaowen Mao, Mohamed Mehana

Machine learning is framed as a way to reduce the cost of simulating cyclic underground hydrogen storage (UHS), where renewable electricity is converted to hydrogen, stored in geological formations, and later recovered. Using 1,000 two-dimensional reservoir simulations, the study finds that autoregressive U-Nets can improve hydrogen-saturation prediction over static models and extrapolate beyond the training horizon, while also exposing challenges in pressure prediction, error accumulation, resolution, and operational generalization.

Key insights

• Autoregressive modeling is well suited to UHS because storage performance depends on repeated operational cycles and evolving subsurface states, but deployment can suffer from accumulated prediction errors after many rollout steps.
• Hydrogen plume behavior may require broad-domain, high-resolution modeling: low viscosity, high diffusivity, geological channels, and cushion gas can produce long-range preferential migration that local refinement alone may miss.
• The results should be interpreted as feasibility evidence rather than a validated operational system, since the experiments use reduced-resolution two-dimensional simulations, fixed operations, limited model comparisons, and do not evaluate the reserved test set.

Summary by alphaXiv.