How much catalyst composition space can be searched from only 70 experiments?

The electrochemical problem

High-entropy electrocatalysts offer a large composition space, but that same flexibility makes systematic experimental optimization expensive. With five variable metals, even a coarse composition grid can produce thousands of candidates.

The study asks whether a relatively small set of measured oxygen-evolution catalysts can be used to build a model accurate enough to choose the next composition worth making.

What the researchers did

The authors measured 70 FeCoCrMnCu layered-double-hydroxide compositions and used the data to train an XGBoost regression model for OER overpotential. The model was then applied across a much larger composition space to identify a promising formulation.

The selected Fe0.15Co0.10Cr0.30Mn0.30Cu0.15 composition was synthesized and electrochemically tested rather than accepted on prediction alone.

What they found

The XGBoost model achieved R² = 0.84 with an RMSE of 9.95 mV. It predicted an overpotential of 261.3 mV for the selected composition, while the experiment measured 270 mV at 10 mA cm−2, a difference of about 3%.

The authors frame the approach as a way to avoid exhaustive testing of more than 10,626 possible compositions within the defined search space.

Why this matters

The interesting point is the leverage gained from a small experimental dataset. Instead of asking machine learning to replace catalyst experiments, the model is used to concentrate experimental effort where the expected payoff is higher.

That is a natural fit for multicomponent electrochemistry, where compositional degrees of freedom grow much faster than the number of samples that can realistically be synthesized and characterized.

One important caution

Seventy compositions remain a small training set, and the reported predictive performance applies to the defined FeCoCrMnCu-LDH family and experimental protocol. A model that interpolates usefully inside this composition family may not transfer to different catalyst structures or testing conditions.

The close agreement for the selected composition is valuable experimental validation, but it is one optimized point rather than a demonstration of uniform accuracy over the entire high-entropy space.