Researchers built an automated, AI-guided platform to search for electrocatalysts that perform the oxygen evolution reaction (OER) in acidic conditions. The OER is the chemical half‑reaction at the anode of proton‑exchange‑membrane water electrolysis (PEMWE), which is important for making hydrogen. Today’s best catalysts for this reaction use iridium or ruthenium, rare metals with tight supply chains. The team aimed to find alternatives that avoid those elements.

The platform combined combinatorial sputter synthesis (a way to make many different thin films), rapid high‑throughput screening, machine learning (ML) models that connect composition to properties, and adaptive multi‑objective optimization. It also used context‑aware large language model (LLM) reasoning for guidance. The loop was human‑supervised but more than 90% automated. Over the campaign the system tested 2,942 different catalysts across 53 material systems and 26 elements, and a sequential learning agent chose new candidates as data accumulated.

The search turned up palladium‑oxide based complex oxides that do not contain iridium or ruthenium. Two notable examples were InMnPdOx (indium‑manganese‑palladium oxide) and NiTaPdOx (nickel‑tantalum‑palladium oxide). In long‑term tests at 10 mA cm‑2 in 1 M H2SO4, NiTaPdOx ran at a lower overpotential than plain PdOx but both eventually exceeded an overpotential of 0.5 V: PdOx at about 200 hours and NiTaPdOx at about 470 hours. InMnPdOx showed both improved overpotential and a large increase in operational stability, keeping the overpotential below 0.5 V for more than 1,000 hours.

Characterization linked this improved performance to the role of the added elements. The additives appear to promote a needle‑like nanostructure that forms while the catalyst is operating. That nanostructure is associated with activity, and the additives also help stabilize palladium against corrosion under the harsh, low‑pH conditions of the test. The paper reports retrospective benchmarking in which the sequential learning agent advanced the activity–stability trade‑off faster than a fixed Bayesian optimization policy or simple in‑context LLM selection.