A research team from the Shanghai Institute of Ceramics, Chinese Academy of Sciences (SICCAS), in collaboration with Beihang University, has developed an integrated strategy that uses artificial intelligence to identify active sites in high-entropy catalysts and automatically validate the predictions in the laboratory. The findings were published in Science Advances.

The team focused on high-entropy CoOOH catalysts for the oxygen evolution reaction (OER)—a key process in water electrolysis for producing clean hydrogen fuel. High-entropy catalysts are composed of multiple transition metals, offering a tunable platform for performance optimization. However, their complexity creates an enormous number of possible atomic configurations, making it difficult to identify which specific sites are responsible for catalytic activity.

To address this challenge, the researchers constructed a computational dataset containing 4,822 structures and developed an attention-enhanced AI model. The model simultaneously predicts two key properties: OER overpotential (a measure of catalytic efficiency) and doping formation energy (a measure of how easily different metal atoms can be incorporated into the catalyst). The model achieved mean absolute errors of 4.5 mV for overpotential and 3.6 meV per atom for formation energy, outperforming conventional methods under small-data conditions.

Using this model, the team screened 17,500 candidate structures and identified eight catalysts predicted to combine high activity with structural stability. These candidates were synthesized and tested using the DREAMLab intelligent laboratory at SICCAS, an automated system for material preparation and electrochemical characterization. Among the experimentally validated catalysts, TiFeNiZn-CoOOH exhibited the highest OER activity, requiring an overpotential of 263 mV at 100 mA cm⁻²—93 mV lower than undoped CoOOH.

To interpret the model's predictions, the researchers analyzed feature importance and examined catalytic mechanisms across more than five million predicted structures. They found that zinc consistently showed the highest probability of occupying active sites, and the [CoNiZn] local coordination environment corresponded to the lowest OER overpotential. Electronic-structure analysis revealed that the incorporation of zinc shifts oxygen orbitals toward the Fermi level, enhancing the availability of electronic states for the reaction and reducing the energy barrier.

This work establishes a closed-loop framework that connects deep learning, active-site identification, automated synthesis, electrochemical validation, and mechanistic interpretation. By linking atomistic prediction with automated experimental verification, the strategy provides a transferable route for the rational design of high-entropy catalysts and accelerated discovery of energy-conversion materials.

Contact: Jianjun Liu / Nian Ran

Shanghai Institute of Ceramics, Chinese Academy of Sciences

E-mail: jliu@mail.sic.ac.cn,rannian@mail.sic.ac.cn

Published online: February 13, 2026

Figure1

Figure 1. Schematic diagram of the workflow for high-entropy CoOOH catalyst design. The workflow integrates multiobjective transfer learning based on EquiformerV2 and the Post-Att Adapter, materials and active-site search, interpretability-assisted mechanism discovery, and automated laboratory validation.