As the world moves toward clean energy, hydrogen is gaining momentum as a promising alternative to fossil fuels. It’s versatile, renewable, and—if handled right—can power everything from industrial plants to cars. But there’s one big question: Where do we store all that hydrogen until we need it?
Enter Underground Hydrogen Storage (UHS)—a method that stores hydrogen deep underground in old gas reservoirs, salt caverns, or aquifers. It’s an elegant solution to the energy storage puzzle. But like most good ideas, it has its challenges.
The Hidden Problem: Cushion Gases and Contamination
When we inject hydrogen underground, we don’t do it alone. We also inject “cushion gases”—gases like nitrogen, methane, or carbon dioxide that help maintain pressure and improve flow.
But here’s the catch: hydrogen and cushion gases mix, and that mixing (called dispersion) can contaminate the hydrogen. That means:
- Lower hydrogen purity
- Higher purification costs
- Less efficient energy systems
To manage this better, scientists need to know exactly how much hydrogen mixes with cushion gases—measured by something called the dispersion coefficient (KL).
Why Traditional Methods Fall Short
Historically, scientists have used lab experiments or computer simulations to estimate KL. But those methods are:
- Expensive
- Time-consuming
- Limited in accuracy, especially under the real, dynamic conditions of underground reservoirs
This is where machine learning (ML) comes in.
A New Era: Using Machine Learning to Predict Hydrogen Behavior
A team of researchers from Persian Gulf University has found a way to use machine learning to solve this dispersion challenge. Their idea: train ML models using real experimental data, then use those models to predict hydrogen dispersion more accurately, faster, and at a lower cost.
The Study
- They used data from experiments with sandstone rock cores.
- Conditions mimicked underground storage (high pressure and temperatures).
- Gases tested: hydrogen + nitrogen (N₂), methane (CH₄), and carbon dioxide (CO₂).
- ML models were trained to predict how hydrogen mixes (KL) based on pressure and flow velocity.
Which Machine Learning Model Won?
The researchers tested several ML models, including:
- Random Forest (RF)
- Artificial Neural Networks (ANNs)
- Support Vector Machines (SVMs)
- Linear Regression
- Bayesian Regression
- Least Squares Boosting
✅ Winner: Random Forest
- R² of 0.9965 on test data (near-perfect prediction)
- Lowest error among all methods
- Required no expensive new experiments
Bonus Insight: CO₂ Is the Most Effective Cushion Gas
One key finding: among the gases tested, carbon dioxide (CO₂) caused the least hydrogen dispersion. That makes it a strong candidate for use in future underground hydrogen storage systems.
Why CO₂?
- It’s denser than hydrogen
- It doesn’t mix easily
- It helps maintain system stability
Why This Matters
This research has real-world implications for the clean energy transition. Using machine learning:
- Saves time and cost compared to traditional lab methods
- Improves predictions for hydrogen purity
- Supports better decision-making in designing underground storage
It also opens the door to real-time monitoring systems that adjust injection and withdrawal based on predictive analytics—making underground hydrogen storage smarter and cleaner.
Looking Ahead
While the results are promising, the researchers acknowledge a few limitations:
- Their model is based on lab data—not full-scale field operations
- Only pressure and velocity were used as variables (future models should include more)
- Geological complexity and time-dependent effects still need more study
Future work will involve testing these models in the field, adding more physical variables, and refining the algorithms to match the messy, real-world conditions of underground reservoirs.
Conclusion
By combining traditional energy engineering with cutting-edge machine learning, this study offers a practical path forward for storing hydrogen underground—safely, efficiently, and affordably. As we race toward a sustainable energy future, innovations like this will be essential.
Smart energy needs smart tools—and machine learning is proving to be one of the smartest yet.
References
- Akbari, A., Maleki, M., Kazemzadeh, Y., & Ranjbar, A. (2025). Calculation of hydrogen dispersion in cushion gases using machine learning. Scientific Reports, 15, Article 13718. https://doi.org/10.1038/s41598-025-98613-9
- Kobeissi, S., Ling, N. N., Yang, K., May, E. F., & Johns, M. L. (2024). Dispersion of hydrogen in different potential cushion gases. International Journal of Hydrogen Energy, 60, 940–948. https://doi.org/10.1016/j.ijhydene.2024.02.151
- Jahanbakhsh, A. et al. (2024). Underground hydrogen storage: A UK perspective. Renewable and Sustainable Energy Reviews, 189, 114001. https://doi.org/10.1016/j.rser.2023.114001
- Maleki, M. et al. (2024). Investigation of wettability and IFT alteration during hydrogen storage using machine learning. Heliyon, 10(19), e38679. https://doi.org/10.1016/j.heliyon.2024.e38679
- Muhammed, N. S. et al. (2023). Hydrogen storage in depleted gas reservoirs: A comprehensive review. Fuel, 337, 127032. https://doi.org/10.1016/j.fuel.2022.127032
- Dąbrowski, W. (2024). Hydrogen and methane dispersion in rock cores. International Journal of Hydrogen Energy.
- Maniglio, G. et al. (2022). Evaluating the impact of dispersive mixing on UHS systems. Scientific Reports.
- Li, J. et al. (2023). Hybrid probabilistic deep learning for hydrogen plume prediction. International Journal of Hydrogen Energy.
