We demonstrate how the integration of machine learning (ML) and density functional theory (DFT) approaches accelerates the discovery of advanced intermetallic alloys for solid-state hydrogen storage applications. By developing data-distribution-imbalance-aware ML models and scalable machine learned interatomic potentials, we efficiently screened vast compositional spaces and identified Pareto-optimal materials with improved gravimetric hydrogen storage capacity and thermodynamic performance. The approach not only yielded accurate predictions aligned with experimental observations but also provided interpretable chemical insights, including empirical substitution. This work underscores the potential of computational and data driven design in addressing critical materials challenges in the hydrogen economy.

The global push towards net-zero carbon emissions has positioned the hydrogen economy at the forefront of the energy transition. Among the various hydrogen storage methods, solid-state storage, where hydrogen is stored within materials via physical or chemical interactions, offers superior safety and capacity compared to high-pressure gas storage. Intermetallic alloys, a well-established class of solid-state hydrogen storage materials, exhibit desirable cyclability and kinetics required for practical applications. However, their low gravimetric hydrogen capacity remains a critical limitation. To address this, we employ machine learning (ML) approaches to accelerate the discovery of low-cost intermetallic alloys with enhanced hydrogen storage properties. Complementary density functional theory (DFT) calculations are used to probe underlying chemical mechanisms of the interactions between hydrogen and intermetallic alloys and generate accurate training data for machine-learned interatomic potentials which are needed for high throughput compositional and configurational samplings.

Materials discovery typically seeks candidates with exceptional properties, which are rare and sparsely represented in existing datasets. Conventional ML algorithms often perform poorly in these underrepresented regions due to their bias towards the data-rich majority. To overcome this, we curated literature data on intermetallic alloys and trained ML models specifically designed to handle imbalanced data distributions. These models showed improved predictive performance for extreme values and enabled the identification of Pareto-optimal compositions for solid-state hydrogen storage and metal hydride compressor applications.

Beyond predictions, our models provided valuable insights into underlying chemical trends. By analysing feature importance and patterns within the surrogate ML models, we derived empirical design rules for elemental substitution in various alloy systems. The ML predicted hydrogen storage properties showed excellent agreement with experimental measurements, demonstrating the validity of our ML-guided approach to compositional design and optimisation.

To support accurate prediction of key thermodynamic properties—specifically, hydrogen pressure-composition-temperature (PCT) isotherms, we developed accurate machine learning interatomic potentials trained on a diverse DFT database generated using ARCHER2. Simulating PCT behaviour requires sampling numerous configurations, which is computationally prohibitive using DFT alone. Our machine learning interatomic potentials overcome this bottleneck, enabling rapid and reliable estimation of PCT curves. These predictions are being validated through targeted DFT simulations. Moreover, we used DFT to investigate selected compositions in depth, uncovering how specific metal substitutions influence hydrogen storage behaviour. This led to the identification of key factors governing thermodynamic performance across multiple intermetallic systems.

This study demonstrates the power of combining machine learning with first-principles simulations to accelerate the discovery and optimisation of intermetallic hydrogen storage materials. By addressing key challenges such as data imbalance and the high computational cost of thermodynamic predictions, our approach enables efficient identification of promising alloy compositions and provides interpretable chemical insights.

Outcomes:

• New knowledge: Machine learning revealed hydrogen absorption plateau pressure trend with respect to metal substitutions in intermetallic alloy systems.
• New capability: An automated pipeline integrating composition screening, structure optimisation and pressure-composition-temperature prediction using machine learning interatomic potential
• New collaborations: Partnerships established with international leading research groups in USA, France and Japan.
• New industrial link to Hyundai Motor Company.

References:

[1] ACS Appl. Energy Mater. 2025, 8, 492

[2] Acta Mater. 2024, 276, 120086

[3] J. Phys. Chem. Lett. 2024, 15, 1500

[4] Int. J. Hydrogen Energy 2023, 48, 13227

[5] J. Mater. Chem. A 2023, 11, 15878

[6] Chem. Mater. 2021, 33, 4067

[7] J. Phys. Chem. Lett. 2019, 11, 40.

Collaborators:

• University of Nottingham (David Grant, Martin Dornheim)
• Sandia National Laboratories (Matthew Witman, Vitalie Stavila)
• Institut de Chimie et des Matériaux Paris-Est, ICMPE – CNRS (Claudia Zlotea, Junxian Zhang)
• Uppsala University (Martin Sahlberg)

Funding:

• EPSRC £973k, Sep 2021 – Feb 2025
• EPSRC £314k, Feb 2025 – Jan 2028
• Hyundai Motor Company

Contact:

Dr. Sanliang Ling (Sanliang.Ling@nottingham.ac.uk), Faculty of Engineering, University of Nottingham