In brief

QuestionCan ML screening remain fast while respecting the potential-dependent and ensemble nature of real electrocatalysts?
ResultThe model screened 22,599 structures, found 3,431 with predicted ORR limiting potentials above 0.8 V, and identified only eight of 729 metal pairs with more than 80% active configurations.
Why it mattersThe study changes the screening target from one ideal structure to a reaction-condition-aware distribution of structures, which is closer to what synthesis actually produces.
Reaction-Condition- and Ensemble-Aware Screening of Dual-Atom ORR Catalysts via Physics-Informed Machine Learning
Prajeet Oza et al. · ACS Catalysis · Published 9 September 2026
DOI: 10.1021/acscatal.6c04252

The electrochemical problem

High-throughput electrocatalyst screening often evaluates a clean, single structure and assigns one activity value to it. Real dual-atom catalysts are less tidy: applied potential can change the preferred surface state, and synthesis can produce multiple local atomic configurations for the same metal pair.

A useful screening method therefore has to be fast enough for thousands of structures while still representing electrochemical operating conditions and configurational diversity.

What the researchers did

The authors combined high-throughput DFT with an equivariant transformer graph neural network. The design space contains 22,599 M1/M2-N-C structures formed from 729 transition-metal pairs and 31 structural configurations.

Instead of directly learning the final limiting potential, the model predicts adsorption energies. These values feed voltage-dependent ab initio thermodynamic phase diagrams, which determine realistic hydroxylation states and ORR limiting potentials. The authors then rank each metal pair by the fraction of its configurations that are catalytically active.

What they found

After iterative data refinement, the reported mean absolute error reached 0.015 eV for adsorption energies and 0.022 V for derived ORR limiting potentials. Screening identified 3,431 structures with predicted limiting potentials above 0.8 V.

Only eight of 729 metal pairs had more than 80% active configurations. Co appeared in seven of those eight pairs, including Co/Cr, Co/Ag, Co/Ru, Co/Ir, and Co/Zn. Additional DFT calculations supported the top-ranked predictions and the framework reproduced several experimental trends from related systems.

Why this matters

The strongest idea is the definition of the target. Experimentalists can usually control the precursor composition more easily than one exact dual-atom geometry. Ranking a metal pair by the fraction of likely active configurations is therefore more actionable than selecting a single perfect structure.

The paper also shows how physics can make ML more data-efficient: learning intermediate adsorption energies provides richer training information than trying to learn the final limiting potential directly.

One important caution

The top candidates are still predictions. Agreement with existing trends and follow-up DFT calculations are useful validation, but the newly proposed metal pairs are not synthesized and tested in this study.

The phase diagrams also represent a defined set of surface states and thermodynamic assumptions. They improve realism substantially without reproducing every kinetic, reconstruction, transport, or durability process in an operating fuel-cell cathode.

Primary source

Prajeet Oza et al. Reaction-Condition- and Ensemble-Aware Screening of Dual-Atom ORR Catalysts via Physics-Informed Machine Learning. ACS Catalysis (2026). DOI: 10.1021/acscatal.6c04252.

Research-note facts and quantitative values were checked against the publisher record linked above. The “Why this matters” and “One important caution” sections are EchemAI editorial interpretation.