Can a pretrained machine-learning potential find a practical noble-metal-free OER catalyst in a large atomic configuration space?

The electrochemical problem

Oxygen evolution is a major efficiency bottleneck in water electrolysis. Noble-metal catalysts can be highly active, but cost and scarcity motivate the search for alternatives that remain effective at device-relevant current densities.

The materials space is large even before synthesis begins: changing the host oxyhydroxide, the incorporated element, and the atomic site creates thousands of possible structures. The central question is whether machine learning can reduce that search to experimentally actionable candidates.

What the researchers did

The study used a pretrained EquiformerV2 model to accelerate adsorption-energy evaluation outside the original training domain. The candidate space contained 3,976 single-atom-incorporated metal oxyhydroxide configurations.

The screening nominated atomically dispersed tungsten in NiFe oxyhydroxide, W1–NiFeOOH. The material was then prepared by cyclic electrodeposition and evaluated in alkaline exchange-membrane water electrolysis. In situ spectroscopy and DFT were used separately to investigate why tungsten changes the catalytic behavior.

What they found

W1–NiFeOOH reached 13.1 A cm−2 at a cell voltage of 2.0 V at 80 °C in AEM water electrolysis using commercial membranes. Long-term testing showed stable operation for more than 500 h at 1 A cm−2 at 60 °C.

The mechanistic analysis suggests that subsurface tungsten acts as a promoter rather than simply replacing a noble-metal active site: it redistributes electronic structure around neighboring Ni–O–Fe edge sites and facilitates deprotonation and transformation toward the active γ phase.

Why this matters

This paper is a strong example of machine learning being useful because it changes what can be screened before expensive experiments are performed. The ML model is not the end product. It is a filter that generates a concrete experimental candidate.

The subsequent synthesis, membrane-electrode-assembly testing, spectroscopy, and DFT analysis give the prediction an electrochemical context. That full chain is much more informative than a screening score alone.

One important caution

The impressive current density and durability belong to a specific catalyst, cell construction, temperature, electrolyte, and membrane configuration. They should not be interpreted as universal figures for all W-containing oxyhydroxides or all electrolyzers.

Likewise, the mechanistic conclusion comes from spectroscopy and first-principles calculations. The machine-learning model accelerated candidate screening; it did not by itself establish the catalytic mechanism.