A Flexible Theory for Catalysis: Learning Alkaline Oxygen Reduction on Complex Solid Solutions

A New Era of Catalyst Design: Machine Learning Meets High-Entropy Alloys

Published in: Angewandte Chemie International Edition (2023)

Authors: Christian M. Clausen, Olga A. Krysiak, Lars Banko, et al.

DOI: 10.1002/anie.202307187

As the demand for efficient, low-cost energy solutions rises, scientists are looking beyond traditional materials to find next-generation catalysts. This study presents a powerful combination of machine learning and high-entropy alloy (HEA) chemistry to predict and understand the behavior of electrocatalysts in the alkaline oxygen reduction reaction (ORR).

Exploring a Complex Composition Space

The research focuses on the compositionally complex solid solution (CCSS) system Ag–Pd–Pt–Ru. Using combinatorial co-sputtering, the team created thin-film material libraries representing 1582 unique alloy compositions. These were analyzed using high-throughput electrochemical testing in alkaline conditions.

Adsorption Energy as the Key Descriptor

Rather than relying solely on alloy composition, the researchers developed a graph neural network (GNN) to predict adsorption energy distributions for key ORR intermediates (*OH and *O). These distributions were then used as descriptors in a theory-derived catalytic model, resulting in highly accurate predictions of ORR activity—with a mean absolute error of just 0.042 mA/cm².

Bridging Theory and Reality

Notably, discrepancies between experimental and predicted results provided insights into how real-world surfaces may differ from their idealized compositions. These deviations hint at changes in surface composition during operation—possibly due to oxidation or segregation—which traditional models often overlook.

Key Takeaways

• Machine learning models, when trained on DFT data, can accurately predict catalytic activity across complex alloy spaces.

• Adsorption energy distributions are better descriptors than simple composition metrics.

• The approach enables extrapolation to unexplored compositions, offering a path to AI-guided catalyst discovery.

• Findings suggest real catalytic surfaces often evolve during reactions, highlighting the need for dynamic models and surface-sensitive characterization.

Conclusion

This work pushes the boundaries of how we understand and design catalytic materials. By integrating machine learning with physical theory and high-throughput experimentation, the study offers a scalable and predictive framework for the discovery of next-gen electrocatalysts—especially within the high-dimensional spaces of HEAs.