Smarter search for fuel-cell catalysts using machine learning - isct.ac.jp

Proton exchange membrane fuel cells (PEMFCs) are a promising clean energy technology that can generate electricity by combining hydrogen and oxygen into water. However, their performance depends heavily on a chemical step known as the oxygen reduction reaction (ORR), which requires an efficient catalyst to proceed at practical rates. Although platinum (Pt) remains the standard ORR catalyst in PEMFCs due to its remarkable electrochemical properties, its high cost and scarcity are barriers to large-scale adoption. As a result, researchers have turned to platinum-based alloys as less expensive alternatives that still maintain strong catalytic performance.

Designing these alloy-based catalysts, however, is far from straightforward. The number of possible atomic arrangements in alloy materials is enormous, making it impractical to test every candidate through experiments or computational methods like density functional theory. At the same time, catalysts must satisfy more than one requirement; they need to be highly reactive for ORR, but also stable under real operating conditions. Most machine learning-based approaches address these properties separately and thus lack the ability to propose atomic structures that fulfill both criteria simultaneously. How can we search for suitable alloy designs more efficiently?

In a recent study, Associate Professor Atsushi Ishikawa of the School of Environment and Society at Institute of Science Tokyo, Japan, together with graduate student Taishiro Wakamiya, developed a new strategy to address this challenge. Their work, published in the journal npj Computational Materials on April 14, 2026, introduces a method that combines atomistic simulations with generative artificial intelligence to design alloy catalysts for the ORR.

The proposed approach hinges on two key tools. The first is a neural network potential (NNP) model—a machine learning model trained on quantum mechanical calculations that can quickly estimate key material properties. The second is a generative model known as a conditional variational autoencoder (CVAE), which can propose new atomic structures based on desired properties. In this case, the model was trained to target both low overpotential (a measure of catalytic activity) and low alloy formation energy (a measure of stability).

The workflow operates as an iterative loop, with the NNP model evaluating the performance of proposed alloys and the CVAE refining them and feeding them back to the NNP stage. Over multiple iterations, this process gradually shifts the alloys toward better-performing arrangements. When applied to Pt–nickel alloys, the method generated structures that met overpotential and formation energy criteria simultaneously. Notably, the model also rediscovered known design principles by itself, such as how platinum-rich surface layers can enhance ORR activity.

The team further demonstrated the versatility of their workflow by extending it to multiple alloy systems, including Pt–titanium and Pt–yttrium. "The present work demonstrates that the combined use of atomistic calculations and the CVAE provides a general computational screening method that can produce new alloy surface structures satisfying both activity and stability criteria from limited initial data," explains Ishikawa.

Beyond fuel-cell catalysts, the researchers believe their framework could have wide-ranging applications. "The newly developed workflow may be applicable to a broad range of materials challenges, including water electrolysis for hydrogen production, battery electrode materials, and catalysts for chemical processes," concludes Ishikawa.

By enabling faster and more targeted exploration of complex material spaces, this work could help accelerate the development of sustainable energy technologies.

Reference

Related articles