Can catalyst knowledge be transferred from simple alloys to high-entropy surfaces instead of learning the hardest chemistry from scratch?

The electrochemical problem

High-entropy alloys contain a huge diversity of local atomic environments. For electrocatalysis, the quantity of interest is often the adsorption energy of intermediates on those local sites, but direct DFT sampling of the entire space is prohibitively expensive.

Training a model only on the small number of available HEA calculations creates a second problem: the data-hungry ML model is being asked to learn the most complex chemistry from the least data.

What the researchers did

The authors use a chemistry-wise transfer strategy. A graph model first learns adsorption behavior from unary and binary intermetallic systems, then transfers to ternary compounds, and finally moves into high-entropy alloys. Active learning is combined with the final HEA stage to focus new calculations on informative structures.

The study uses both in-house DFT adsorption data for selected HEAs and filtered data from the Open Catalyst 2020 dataset.

What they found

The sequential transfer workflow improved adsorption-energy prediction for CO2-reduction intermediates compared with models trained directly on the limited HEA dataset. The authors also use t-SNE analysis to argue that chemically relevant features are being transferred across the composition domains.

The result supports a hierarchical learning strategy in which simpler materials supply reusable chemical information before the model tackles the combinatorial HEA space.

Why this matters

In electrochemical materials discovery, data scarcity is often worst exactly where materials are most complex. Transfer learning offers a way to exploit calculations that already exist for simpler systems rather than discarding them as chemically irrelevant.

The approach is conceptually similar to how researchers build intuition: first learn recurring bonding and adsorption patterns in simpler environments, then refine that knowledge for a more complicated material.

One important caution

The target in this study is adsorption-energy prediction, which is an important catalyst descriptor but not a direct measurement of CO2-reduction rate, selectivity, stability, or device performance.

The work therefore demonstrates more efficient computational exploration of HEAs; experimental electrocatalyst discovery remains a subsequent step.