The scarcity problem: iridium is almost gone
Green hydrogen runs into a stubborn problem at the electrolyzer stage: there is not enough iridium.
The anode of a proton exchange membrane (PEM) electrolyzer must drive the oxygen evolution reaction (OER) in strong acid at high potentials, where nearly all metals dissolve. The industry-standard answer is iridium oxide—but iridium is one of Earth's rarest elements, produced almost entirely as a byproduct of platinum mining, with global annual output of only a few tons. Ruthenium is a second-tier option, less stable and equally scarce.
On September 25, 2026, Lila Sciences announced a batch of results. Its AI closed-loop laboratory proposed, synthesized, and screened 2,942 catalysts in about three months, spanning 53 material systems and 26 elements, and identified 6 palladium-based material families containing neither iridium nor ruthenium. One of the lead candidates, InMnPdOx, maintained an overpotential below 0.5 V long-term in 1 M sulfuric acid at 10 mA/cm², running stably for over 1,000 hours.
The accompanying preprint, submitted to arXiv on September 24, is titled *AI-guided high-throughput discovery of iridium- and ruthenium-free palladium-oxide catalysts for durable acidic oxygen evolution*, arXiv:2609.30133, first author Ken J. Jenewein, with roughly twenty co-authors.
What 1,000 hours means
In days, the comparison becomes clearer:
| Material | Time before crossing the 0.5 V overpotential threshold | |---|---| | Unmodified PdOx | ~200 hours (~8 days) | | NiTaPdOx | ~470 hours (~20 days) | | InMnPdOx | Over 1,000 hours, still below 0.5 V (~42 days) |
Lila states that the optimized metal composition's activity is comparable to industry-standard ruthenium-based catalysts. Note this is a vendor-reported comparison; baseline details belong to the preprint.
Why palladium was considered hopeless
Palladium was long a dead end for acidic OER—its poor performance was an entrenched view in the literature. Lila's approach was doping: not pure palladium, but palladium modified with small amounts of other elements. Adding indium and manganese lifted both activity and stability.
Two quotes worth remembering
John Gregoire, Lila's Chief Autonomous Science Officer, led a group at Caltech that studied such catalysts for twelve years. He said that if a student had proposed this element combination back then, he would have advised them to switch to a more promising direction.
> "This combination defies all traditional wisdom."
Rafael Gómez-Bombarelli, Chief Science Officer of Physical Sciences at Lila and MIT's Paul M. Cook Associate Professor, added that without lab results backing it up, the team would have dismissed the result as a hallucination.
Assessment: These two quotes capture the real core of the story. AI's role here was not "computing faster" but proposing a direction human experts would have vetoed by intuition—and pushing it onto the experimental bench.
How the closed loop runs
Lila began building its AI Science Factory (AISF) system in late 2024, chaining four modules:
- Synthesis → pre-test characterization → acidic OER testing → post-test characterization → AI model → back to synthesis
- This is lab-scale material discovery, not an industrial product; device- and industrial-form durability is still under evaluation
- The preprint has not completed peer review; public data come from Lila's own announcements
- The >90% automation figure retains manual steps
- The 1,000-hour result corresponds specifically to 10 mA/cm² constant current in 1 M sulfuric acid; conclusions must be retested at industrial current densities and in real electrolyzer environments
- Counter-intuitive directions are worth more than speed. Screening in three months is engineering capability; finding a combination experts would have discouraged is the scientific increment
- The supply constraint is real. Iridium's annual output is measured in tons and depends on platinum mining byproducts—a physical ceiling on electrolyzer scale-up
- Closed-loop labs are becoming infrastructure. Lila says the same pipeline will next target ammonia and plastic precursors; competition shifts from "model scores" to "scalable experimental throughput and data feedback"
- For the supply chain: if iridium-free acidic OER routes hold up industrially, palladium-based families could reshape the electrolyzer supply narrative
- Whether the preprint survives peer review and whether third parties independently reproduce InMnPdOx's 1,000-hour curve
- Whether Lila publishes pre/post-cycling characterization, especially palladium corrosion and nanostructure formation
- When device-level data arrives—the only threshold from "material discovery" to "usable component"
- Whether other autonomous labs reproduce the result on the same sub-direction; a single-source claim and a multi-lab replicated result carry very different evidentiary weight
Each cycle starts with the AI analyzing chemical space and proposing new formulas; pipeline output feeds back into the models. The technical stack includes combinatorial sputtering synthesis, high-throughput screening, composition–property machine learning models, adaptive multi-objective optimization, plus LLM reasoning. Internally, Bayesian models reason under uncertainty while language models carry broad context about how the world works, balancing information gain against uncertainty.
Throughput numbers: about 240 samples per week, roughly 17× faster than traditional methods, with over 90% savings in per-sample labor time. Lila claims over 90% automation but explicitly notes human supervision and sample transfer remain—not fully unattended.
Where humans sit in the pipeline
A detail easily missed: electrochemist Ken Jenewein initially reviewed AI-proposed suggestions one by one, rejecting some, then gradually withdrew from review.
Inference: This exit timeline says more about system maturity than any "AI autonomy" marketing line—it implies the human rejection rate eventually fell to a level worth tracking but not worth intervening on every cycle.
Boundaries: this is still a lab result
Assessment: The correct reading today: one more candidate direction added to the lab-scale Pareto frontier. Translating this into "rewriting the green hydrogen cost curve" requires at least two more races: device-level validation and scale-up manufacturing.
Why this matters
What to watch next
Sources
1. Jenewein K. J. et al., *AI-guided high-throughput discovery of iridium- and ruthenium-free palladium-oxide catalysts for durable acidic oxygen evolution*, arXiv:2609.30133, 2026-09-24 2. Lila Sciences results announcement (by Megan Scudellari), 2026-09-25 3. Unite.AI, *Lila Sciences' AI Lab Uncovers Palladium Catalysts for Green Hydrogen*, 2026-09-25 4. AdvanceH2, *AI-Powered Discovery of Palladium Catalysts Marks a Breakthrough in Green Hydrogen Production*, 2026-09-25 5. AGI Hunt daily briefing, 2026-09-26