The particle histories behind a fuel-cell catalyst’s decline

A fuel-cell catalyst can lose useful surface in several ways. Metal can dissolve, particles can separate from their support, or neighbouring particles can join into larger structures. A measurement of declining electrical performance does not identify which history produced it. Nor does a collection of photographs taken at unrelated locations.[1]

Ana Rebeka Kamšek and colleagues have combined electron microscopy with machine learning to preserve those individual histories across a larger sample. Their peer-reviewed paper, published in accepted form by Nature Communications on 10 October, follows the same platinum–cobalt nanoparticles before and after electrochemical treatment.[1]

Returning to the same particles

The material was a commercial catalyst with metal particles averaging about 2–3 nanometres on a carbon support. One dataset contained 209 classified events from six pairs of images after an activation treatment. A larger dataset contained 696 events spanning activation followed by 5,000 further voltage cycles. The experiments used a laboratory electrode arrangement, rather than an operating commercial fuel cell.[1]

An event is not necessarily one particle. A disappearing particle leaves an outline in only the earlier image; a merger can connect several earlier outlines to one later structure. The researchers linked the outlines across images, then used rules based on particle number, size, roundness and nearby changes to assign a category.[1]

The missing-particle ambiguity is consequential. An outline that vanishes cannot, on its own, reveal whether the metal left as a particle or dissolved into the surrounding liquid. The paper explicitly retains that limitation in its classification. Smaller particles were more likely to fall into the disappearance category, while small and irregular particles rarely remained unchanged.[1]

The laboratory setting also affects the balance between mechanisms. In this electrode geometry, dissolved metal can disperse into a comparatively large surrounding electrolyte volume, limiting its local redeposition. The frequencies therefore describe these experiments, rather than supplying a universal breakdown of how fuel-cell catalysts age.[1]

Finding two particles inside one bright patch

Automation begins with a deceptively difficult image problem: particles overlap. Splitting their shared pixels between two objects can underestimate each particle’s size. The team used CeDiRNet, a neural network that predicts directions towards particle centres, together with estimates of radius and circularity. Multiple centres can be located even where the projected outlines overlap.[1]

The detection network was trained with 23 pairs of microscopy images. Its reported relative errors were 6.35% for particle size and 2.19% for circularity. The supplement specifies that those size and shape errors were calculated only for correctly located detections; they do not measure the accuracy of the complete degradation analysis.[1][2]

A separate matching algorithm aligns the earlier and later images and associates particle outlines. The final mechanism assignment uses a decision tree. These are distinct steps: locating a particle correctly does not guarantee that it is linked to the right later object or assigned the right physical explanation.[1]

Automating association missed about 10% of events, particularly in crowded regions. The fully automated analysis retained the two dominant categories, shrinking and unchanged particles, although their order reversed in the larger dataset. Rarer outcomes were harder to preserve, including the distinction between particles merging and metal depositing onto an existing particle.[1]

Why the history matters to useful surface

A simple geometric illustration shows why merging matters even if the metal remains present. Suppose two separate spherical particles, each two nanometres across, merge into one sphere without losing material. Conserving volume makes the new diameter about 2.52 nanometres. Its surface area is about 20.6% smaller than the combined area of the original pair. This is an idealised calculation, not a measured loss of catalytic performance in the study.

Real catalytic performance also depends on surface chemistry, access to reactants and electrical contact. A particle can appear unchanged because it is shielded by carbon or poorly connected electrically. Alternatively, dissolution followed by redeposition can partly restore its original size. The authors offer these as possible explanations for apparently stable particles; a before-and-after outline does not distinguish them all.[1]

The public research archive contains microscopy images, electrochemical measurements and code for detection, association and degradation analysis. Those materials preserve the connection between the observed object and the category assigned to it.[3]

Catalyst designers need particles that remain attached, accessible to reactants and electrically connected. Recording what happens to individual particles gives those requirements a measurable history. It also shows why an apparently intact piece of metal and a useful catalytic surface cannot be treated as interchangeable.

Study credit: Kamšek and colleagues, Nature Communications, CC BY 4.0. This independently written explanation adapts the study’s findings and adds calculated percentages and a geometric illustration.[1][4]

Report an error or suggest a correction