How Machine Learning and Catalysts are Shaping the Future of Clean Energy

Insights and Learnings from the COST Action NanoCatML Workshop: How Computational Modeling is Transforming Catalyst Design.

The key to advancing clean energy technologies—whether converting CO₂ to higher value chemicals or driving hydrogen innovations – lies in developing highly efficient catalysts.

Computational Material Modeling: Why Guess When We Can Simulate?

Catalyst development is no easy feat. Understanding reaction mechanisms, pinpointing active sites, and achieving optimal energy barriers all play a role in designing better materials. While experiments are essential, simulations can guide us in designing catalysts faster and with higher precision.

Yet, the cost of high-accuracy simulations has been a major hurdle. Researchers often face a trade-off: simplify models to save time or invest significant resources in accuracy. Is machine learning (ML) about to break down these barriers?

New Trends in Machine Learning: Hype or Game Changer?

At the COST Action NanoCatML Workshop in Freiburg, Aixelo joined leading experts to explore how ML can accelerate catalyst design. Here are my three key takeaways that could reshape the field:

Takeaway 1: Machine-Learned Interatomic Potentials Have Reached a Tipping Point

Machine-learned interatomic potentials are emerging as one of the most promising solutions for accurate yet computationally efficient simulations. A standout example is the MACE-MP=0 [1] foundation model, trained on a vast dataset of DFT calculations on 150,000 inorganic crystal structures from the Materials Project [2]. The beauty of such models? Once trained, they require only a small amount of additional data – sometimes as little as 50 data points [3] – to achieve high accuracy at a low computational cost for specific applications. As a downside, MACE-MP-0 supports so far only PBE+U as level of theory.

Another powerful approach is active learning, where the machine-learned interatomic potential is refined by iteratively adding structures to the training data that provide the most useful information [4].

Takeaway 2: Bridging the Gap Between Simulations and Experiments

One of the biggest hurdles in catalyst development is connecting theoretical simulations with practical experiments. A compelling example presented by Sandip De from BASF showed how ML is closing this gap. Their team used density functional theory (DFT) simulations to develop 92 key mechanistic insight-derived descriptors for catalyst activity. These descriptors were then used to train ML models that accurately predict CO₂-to-methanol conversion yield, reducing the need for exhaustive lab testing and significantly speeding up development [5].

Takeaway 3: Data and Computational Workflows Are Becoming More Complex

As ML continues to integrate with simulations, managing data and workflows has become increasingly intricate. Previously, computational studies relied on relatively straightforward input files for quantum chemistry software. Now, with ML-driven approaches, researchers must track a host of additional factors: What data was used? Which ML model was trained? How was it trained? What hyperparameters were optimized?

Jan Janssen highlighted the importance of tools like Pyiron [6] (jan-janssen.com) to handle this complexity. Such platforms help ensure transparency, reproducibility, and efficiency in managing computational workflows.

The Future of Materials Science: What’s Next?

For Aixelo and the broader materials science community, these advancements highlight the need to make cutting-edge tools more accessible. Moving forward, we must focus on building research infrastructure that:

• Organizes data in a structured format and connects it to the workflows through which the data was created, ensuring transparency and reproducibility.
• Supports scientists by organizing AI tools and establishing standardized workflows that better integrate virtual simulations with physical experiments.

The future of catalyst design is unfolding before us—are we ready to embrace it?

References

[1] https://arxiv.org/pdf/2401.00096

[2] https://next-gen.materialsproject.org/

[3] https://pubs.rsc.org/en/content/articlelanding/2025/fd/d4fd00107a

[4] https://www.nature.com/articles/s43588-023-00406-5

[5] https://pubs.rsc.org/en/content/articlehtml/2004/jn/d3cy00148b

[6] https://pyiron.org/