An AI-run lab overturns palladium's dead-end label for acidic oxygen evolution: a modified palladium candidate passes 1,000+ hours of acid stability testing

An AI-driven closed-loop laboratory found a palladium-based catalyst direction for acidic oxygen evolution that the field had long excluded; the lead modified-palladium candidate passed more than 1,000 hours of acid stability testing and performed comparably to ruthenium. In commercial electrolyzers, the catalyst able to survive acid while staying active is iridium oxide, with ruthenium as another option but less stable, and both are among Earth's rarest elements; palladium was considered a dead end because it was assumed to underperform in acidic conditions. While screening 2,942 oxides across 53 systems and 26 elements, the platform began synthesizing and testing palladium compositions, and repeated activity tests plus more than 1,000 hours of stability testing confirmed the lead candidate; the company's scientific leads first judged the model to have gone astray, and the experiments overturned that judgment.

The workflow links synthesis, pre-test characterization, testing and post-test characterization into a closed loop, with each cycle's data flowing back to the AI model, which ranks candidate recipes for the next round. Conventional discovery has a graduate student synthesize one composition at a time, then test and iterate, which cannot navigate the combinatorial space at scale. The AI model combines Bayesian models, which reason under uncertainty, with language models that carry broad context about how the world works, and the interplay balances uncertainty against information gain; experiments in Lila's AI Science Factory synthesized 96 catalysts in parallel by physical vapor deposition and ran accelerated OER screening in acid, measuring both activity and stability.

This is a materials science discovery, not a scaled-up commercial product; any OER catalyst must work at industrial scale under intense operating conditions, and the team is evaluating, with AI guidance and lab equipment, how the material behaves in a form factor resembling industrial use while continuing long-term durability testing. The workflow is not yet fully autonomous: in this first science campaign, humans transferred samples between machines, in parallel with the company's development of robots for full autonomous operation.

The team says it is still scratching its heads: this well-known catalyst does not work for OER on its own, and the model found the right ingredients to make it active and stable, without an explanation of the mechanism. The lead candidate's stability data stop at more than 1,000 hours of acid testing, and long-term durability testing continues.