Deep divefrom Episode 7New physics & chemistry19 October 202627 min

The AI Lab That Found Palladium Where No One Looked

A deep dive into one paper: what was known before, what the researchers did, what the numbers say and what they don't.

The paper: AI-guided high-throughput discovery of iridium- and ruthenium-free palladium-oxide catalysts for durable acidic oxygen evolution (Jenewein, Habib Zadeh et al., Lila Sciences, arXiv preprint)

MP3

Chapters

Also linked

Code and data: fl97inc/lila_oer_manuscript

Transcript

The script the two AI voices read from. The times are worked out from the script, not heard, so a line can start a few seconds early or late.

Read the whole episode

Cold open

Host 1You heard the trailer — this is the whole story. Welcome to the Hastrup Review of Science — where artificial intelligence reads real science, so you don't have to pretend you understood the paper. Today, a robot-heavy lab says its software found a palladium-based catalyst that runs for a thousand hours in strong acid without iridium or ruthenium.

Host 2The paper is called AI-guided high-throughput discovery of iridium- and ruthenium-free palladium-oxide catalysts for durable acidic oxygen evolution. It has twenty-two authors, led by Ken Jenewein and Faezeh Habib Zadeh, with John Gregoire and Rafael Gómez-Bombarelli as senior authors.

Host 1And where are they?

Host 2Every author works at Lila Sciences, a company in Cambridge, Massachusetts, which funded the whole study and says it has filed or may file patents on the materials and the platform. It is an arXiv preprint, version one, posted in September twenty twenty-six, so not yet peer reviewed. Where I add textbook background, I will say so.

Host 1The claim, in two sentences.

Host 2A mostly automated lab, steered by machine learning and a language model, tested two thousand nine hundred and forty-two thin-film oxide catalysts and kept landing on palladium oxide with tiny amounts of other metals. The best one, palladium oxide with traces of indium and manganese, stayed below half a volt of overpotential for a thousand hours in acid, while plain palladium oxide crossed that line at about two hundred hours.

Why this field exists

Host 1Start with the basics. Why does anyone care about oxygen coming off an electrode?

Host 2Background, not from this paper. If you want hydrogen without fossil fuels, you split water with electricity: hydrogen forms at one electrode, oxygen at the other. The hydrogen side is the easy half; making oxygen is the slow, energy-hungry half, and it is where most of the extra voltage goes.

Host 1Extra voltage meaning what, exactly?

Host 2Background again. Thermodynamics says splitting water needs one point two three volts, and anything you must apply beyond that to get a useful current is called overpotential, which is wasted as heat. A good catalyst lowers that overpotential, so the whole number in this episode, zero point four five volts and so on, is the penalty, and smaller is better.

Host 1And why acid?

Host 2Background. One of the leading electrolyzer designs, the proton-exchange membrane electrolyzer, puts a thin acidic polymer membrane between the electrodes, which makes it compact, responsive and good at following the ups and downs of wind and solar power. The price is that the oxygen electrode sits in a strongly acidic, strongly oxidizing environment, which dissolves almost every metal oxide you put there.

Host 1Almost every one.

Host 2The paper puts it plainly: in practice the field relies almost exclusively on oxides of iridium or ruthenium, the only materials that deliver enough activity and durability at the same time. Ruthenium oxide is more active but corrodes faster; iridium oxide is the durable workhorse — that last part is standard background.

Host 1So what is the problem with iridium?

Host 2There is not much of it. The paper calls iridium one of the rarest industrially relevant elements, with ruthenium only slightly more abundant, and both come out as by-products of platinum-group mining through a concentrated supply chain that cannot simply ramp up. As a rough background figure, world iridium output is a handful of tonnes a year.

Host 1And electrolyzer plans are measured in gigawatts.

Host 2Right, and the paper cites analyses asking whether iridium itself could cap how fast that buildout goes. Interestingly, the authors are careful to say catalyst cost is not the main issue, since hydrogen cost depends more on efficiency and lifetime than on precious-metal loading; the worry is supply security.

What was known before

Host 1What did people already know about alternatives?

Host 2Background plus the paper's own framing. Earth-abundant oxides, especially cobalt- and manganese-based ones, can do acidic oxygen evolution for a while, and some have been pushed into real cells, but none has matched iridium on durability over the thousands of hours industry needs. The paper says only a few candidates have approached that bar, and none has met it.

Host 1And palladium?

Host 2Palladium is itself a platinum-group metal, and palladium oxide has a reputation in this reaction for being relatively inactive and prone to dissolving. The authors cite earlier dissolution studies that steered people away from it, and only one prior report of metallic palladium working through strain engineering.

Host 1How is discovery normally done?

Host 2By the paper's description, serially: an expert picks a candidate from theory, literature or intuition, someone makes it, someone tests it, repeat. High-throughput methods speed up individual steps, but screening activity takes seconds per sample while durability takes hundreds of hours, so most fast screens never really measure what matters most.

Host 1Where does AI come in?

Host 2Two kinds. One kind trains on large simulated datasets, which the paper notes has expanded the list of plausible crystals enormously, but simulation is weak here because these catalysts rebuild their own surface once the current flows. The other kind is the closed-loop lab, where a model picks the next experiment, sees the result and updates, which the paper says is still rare in electrocatalysis.

Host 1And language models?

Host 2This is the part the authors are pointed about. They cite work showing off-the-shelf language models used as experiment pickers lean on what their training data says and adapt poorly to new results, in one case performing no worse when the true outcomes were replaced with shuffled labels. That is a bad property if the answer lies somewhere the literature did not expect.

The open question

Host 1So what is the open question?

Host 2Two of them, stacked. Can a closed loop optimize activity and durability together in harsh acid, rather than activity alone, and carry its leads all the way to long tests? And when it chooses the next experiment, what should the decision-maker be: a fixed statistical recipe, a language model, or some mix?

Host 1Why does the mix matter?

Host 2Because the different recipes trade off differently. A recipe that polishes the best-known region is efficient but never leaves home, while one that explores finds surprises but wastes runs; the paper's bet is a system that switches between them as evidence builds up.

What the paper did

Host 1Walk me through the actual lab. How do they make the catalysts?

Host 2By sputtering: inside a vacuum chamber, three or four metal targets are bombarded so atoms fly off and land on small two-by-two-centimetre chips, with oxygen in the gas so the film grows as an oxide. With the chips held still, each spot gets a different mixture, so one run produces a gradient of compositions.

Host 1How many per run?

Host 2Each batch holds thirty-two chips, each with three composition regions that get tested separately. They are split into two fixtures of sixteen chips, so forty-eight catalysts per fixture.

Host 1From which elements?

Host 2Twenty-six of them, including cobalt, nickel, manganese, iron, tantalum, zirconium, indium, palladium and tungsten, but notably no iridium, ruthenium or platinum. Some films are then baked in air, and the software picks the annealing temperature.

Host 1Then what?

Host 2Each film gets a camera inspection and X-ray fluorescence, which reads composition and how much material is on the chip. If X-ray fluorescence cannot resolve the peaks, the software can call in electron-microscope-based analysis instead.

Host 1And the electrochemistry?

Host 2A scanning cell, a little probe that seals onto one spot at a time, in one molar sulfuric acid at room temperature. It holds ten milliamps per square centimetre for ten minutes and records the voltage, then runs four fast voltage sweeps as a stress test, then holds the current again for a minute.

Host 1So activity is the voltage, and stability is the change.

Host 2Stability is two things: how much the overpotential shifts after the stress test, and how much metal was lost, by comparing X-ray fluorescence before and after. That metal loss, in micromoles per square centimetre, is the durability axis for the whole campaign.

Host 1And then the promising ones graduate?

Host 2Yes, to H-cell tests, a classic two-chamber glass cell, first up to four hours, then for leads from four up to a thousand hours, still at ten milliamps per square centimetre. The long runs are built from twelve-hour chunks, with an impedance measurement between chunks so they can correct for resistance losses.

Host 1How automated is this, really?

Host 2They say over ninety percent of the lab work is automated; humans move sample fixtures between stations and start instruments. One full cycle from sputtering to post-test analysis takes about twenty-two hours, and the platform averaged about two hundred and forty catalysts per week.

Host 1Now the brain. What decides the next experiment?

Host 2Four layers. First, surrogate models: machine-learning models that predict overpotential and metal loss from composition-based descriptors and processing conditions, retrained every round, with an automated step that picks whichever model type cross-validates best at that moment.

Host 1Why switch models?

Host 2Because as the explored space grows, a different model may generalize better. Uncertainty comes from bootstrap ensembles, meaning many models trained on resampled data, and the spread between them says how unsure the prediction is.

Host 1Second layer?

Host 2A portfolio of acquisition strategies, which are rules for turning predictions into a shopping list. One is expected hypervolume improvement, the standard choice for improving two objectives at once; others are Thompson sampling, some random exploration, and a lookahead rule that asks how much a candidate would improve future decisions, not just whether it looks good now.

Host 1Hypervolume. Translate.

Host 2Picture a chart with overpotential on one axis and metal loss on the other, both of which you want low. The best trade-offs form a curve called the Pareto front, and the hypervolume is the area that front dominates, so pushing the front outward makes the number go up.

Host 1And then there is active search.

Host 2Active search is a different goal: hunting for rare, high-performing pockets the current front does not contain yet. It matters, because that is the mode in which the live campaign first reached the palladium family.

Host 1Third layer, the language model.

Host 2The language model does not predict performance. It chooses which acquisition strategy to use each round, tunes its settings, picks among the top-ranked candidates and writes a rationale, drawing on a running notebook filled only with observations from the campaign itself.

Host 1And the fourth layer is people?

Host 2Human review, especially at early candidate approval. So it is human-supervised, not autonomous in the strong sense, and the authors say so.

Host 1Which language model?

Host 2For the benchmark comparison, Claude Opus four point six from Anthropic, described as the frontier model at the time. In the interest of disclosure, this show's script is written by a Claude model too, and the authors also used Opus four point six to polish their prose.

A word from our sponsor

Host 1Before the results, a word from our sponsor.

Host 2Does your beard drift slowly upward over two hundred hours? Is it losing material at the edges? The Hastrup Review of Science is brought to you by beard oil.

Host 1I do not have a beard.

Host 2Beard oil is applied in dilute amounts. A few drops across a dense, needle-like structure, and the structure holds.

Host 1Has anyone measured the retention?

Host 2Not by inductively coupled plasma mass spectrometry. Beard oil is available now, and our referral link is in the show notes.

Results

Host 1So, the results. What did two thousand nine hundred and forty-two catalysts look like?

Host 2They span fifty-three material systems and twenty-six elements. For the trade-off chart they kept the one thousand one hundred and twenty-five with overpotential below zero point nine volts and a metal loss large enough to measure reliably, and most of those fall short on either activity or durability.

Host 1And palladium kept winning.

Host 2Palladium-containing compositions consistently sat on the best trade-off front, even though only two hundred and fifty-nine of those one thousand one hundred and twenty-five contained palladium. Two systems stood out, both palladium-rich and not annealed, one with indium and manganese added, one with nickel and tantalum.

Host 1Numbers for those two in screening?

Host 2Overpotentials below zero point four six volts at ten milliamps per square centimetre, with metal loss of zero point three to zero point seven micromoles per square centimetre, meaning they kept roughly sixty to seventy-five percent of their material through screening. The authors say palladium-rich oxides stabilized by dilute additives had not been described as a family for this reaction, so they call it unexpected.

Host 1Anything without platinum-group metals?

Host 2Yes, and it is a useful negative. The best were cobalt-rich with zirconium, one with indium and one with iron, reaching about zero point five five volts, which fits the known picture that cobalt oxides are among the few cheap materials that work in acid at all.

Host 1But?

Host 2Both failed the medium-term test, lasting under four hours. The authors call that space a meaningful improvement with a clear durability ceiling.

Host 1Head to head in screening. How close are the palladium leads to ruthenium oxide?

Host 2They made benchmark films of ruthenium oxide and plain palladium oxide on the same platform. In the first ten-minute hold, the indium-manganese palladium oxide sat at zero point four five volts, within forty millivolts of ruthenium oxide at zero point four one, and about ninety millivolts better than the best cobalt material; those are not resistance-corrected.

Host 1And the stress test?

Host 2After the four voltage sweeps, both palladium leads and plain palladium oxide got slightly better, not worse. Since plain palladium oxide did it too, the authors attribute that activation to palladium oxide surfaces in general, not to the additives, while the cobalt-iron-zirconium material degraded.

Host 1Now, the composition. How much indium and manganese is in there?

Host 2Very little. They measured it properly with mass spectrometry on dissolved films and found ninety-nine point nine one percent palladium by atom count, with zero point zero six percent manganese and zero point zero three percent indium.

Host 1So it is palladium oxide with a rumour of indium.

Host 2That is a fair description of the composition, yes. The nickel-tantalum version is less dilute: two percent nickel, two percent tantalum, ninety-six percent palladium.

Host 1They screened gradients. Did the uniform films behave the same?

Host 2They re-sputtered all four as uniform films on the same platinum-coated silicon substrate, and the screening overpotentials matched the gradient versions within fifteen millivolts. That is a decent reproducibility check between the discovery format and the validation format.

Host 1Now the long test.

Host 2Three-electrode H-cells, one molar sulfuric acid, room temperature, ten milliamps per square centimetre, in twelve-hour chunks. Plain palladium oxide drifts slowly upward over its eight hundred and twenty-eight hours and crosses half a volt of overpotential at about two hundred hours.

Host 1Ruthenium oxide?

Host 2Similar early drift, then a sharper rise at six hundred and fifteen hours, which coincided with the film peeling off the substrate. So that failure is partly the film coming loose, not purely the chemistry.

Host 1Nickel-tantalum?

Host 2Steady for roughly five hundred hours, below plain palladium oxide, then a sharp upturn near five hundred and twenty hours, crossing half a volt at about four hundred and seventy.

Host 1And indium-manganese.

Host 2Below half a volt for the full thousand hours, and it is the most stable trace of the four. A separate film on a porous titanium support, the kind used in real electrolyzers, was also run long-term, partly at a lower current of seven and a half milliamps per square centimetre.

Host 1Metal retention?

Host 2By X-ray fluorescence, the indium-manganese film kept ninety-six percent of its palladium after a thousand hours. Nickel-tantalum kept eighty-eight percent after five hundred and seventy-five hours, plain palladium oxide eighty-seven percent after eight hundred and twenty-eight, and ruthenium oxide fifty-nine percent of its ruthenium after six hundred and sixty-one, measured only where film remained.

Host 1How many long runs is this?

Host 2The headline thousand-hour curves are one cell per material. Replicates exist in triplicate, but only checked at twenty-four and two hundred and sixty-four hours, so the full thousand hours rests on single runs.

Host 1Why would almost nothing added change durability?

Host 2This is where the microscopy comes in. Before testing, the indium-manganese film and plain palladium oxide look similar, both crystalline palladium oxide, though the doped one has smaller crystals, about seven nanometres against ten, and a lattice shrunk very slightly, which they attribute to manganese.

Host 1And after a thousand hours?

Host 2They diverge. Plain palladium oxide clumps into aggregated particles and loses its palladium oxide diffraction peaks, while the indium-manganese film turns into three layers: a needle-like surface network, a porous branched layer under it, with crystal domains no bigger than two nanometres, and a bottom layer that still looks like the original film.

Host 1Is the surface chemistry different too?

Host 2Apparently not. X-ray photoelectron spectroscopy finds similar palladium oxidation states on the surfaces of both after testing, which the authors read as structure, not surface chemistry, governing the difference.

Host 1Does nickel-tantalum also grow needles?

Host 2Yes, a dense needle-like structure, but it failed sooner. The authors' conclusion is modest: the additives seem to steer how the catalyst restructures, and durability over hundreds of hours appears to hinge on subtle differences in that nanostructure.

Host 1Now the AI claims. Did the agent actually beat the alternatives?

Host 2They tested that retrospectively, replaying selection on the finished dataset: random search, fixed-policy Bayesian optimization, fixed active search, and Claude choosing from results in context with no surrogate model. Averaged over ten runs, the agent pushed the trade-off front out fastest, Bayesian optimization did about as well as random, and the language-model-only picker did worse than random.

Host 1Worse than random. Ouch, for the family.

Host 2Noted. The authors connect the poor Bayesian optimization result to a known difficulty, a search space mixing categories, like which elements, with continuous numbers, like fractions and temperatures, which standard Gaussian-process methods handle badly.

Host 1And the second benchmark?

Host 2That one asks a narrower question: starting from scratch, how many three-element systems does each strategy need before it picks one containing palladium with ninety percent probability? Every strategy queries the same machine-learning model trained on all the campaign data, which stands in for the lab.

Host 1And?

Host 2The agent needed eleven systems, active search ten, Bayesian optimization and random search eighteen each. The language model never got there within fifty: with feedback it found palladium in five of twenty runs, only in the last few steps, and without feedback in none of twenty.

Host 1How hard is the target, by chance alone?

Host 2Two hundred and seventy-six of the two thousand three hundred possible three-element systems contain palladium, about twelve percent. Random search found one within budget in nine thousand nine hundred and eighty-seven of ten thousand runs, which tells you how strongly the language model avoided it.

Why it matters

Host 1Step back. What is the principle here?

Host 2For acidic oxygen evolution, durability, not activity, is the scarce property, and it seems to depend on how a material rebuilds itself under current, something simulations cannot yet predict. If you cannot compute it, you have to measure it, and then the question becomes how to choose measurements, which is what this platform is for.

Host 1And the AI principle?

Host 2That a language model is more useful here as a manager of statistical tools than as the tool itself. On its own, an off-the-shelf model's prior that palladium oxide is a poor oxygen catalyst kept it away, while uncertainty-driven exploration went there because the data, not the literature, said so.

Host 1What changes for whom?

Host 2For electrolyzer makers, nothing yet, but the paper's argument is that palladium could become a second source: it is still a platinum-group metal, but its annual supply is more than ten times iridium's, its main use in car exhaust catalysts is declining with electrification, and it is recycled through more varied channels. The authors say it relaxes the supply constraint rather than removing it.

Host 1What does it not mean?

Host 2It does not mean iridium is replaced, or that a palladium anode is ready for an industrial electrolyzer. It does not show an AI discovered chemistry no human could find; it shows a well-designed exploration policy got to an unfashionable region faster than the alternatives they tested, in a retrospective comparison.

Caveats

Host 1Now the caveats. Start with the conditions.

Host 2Room temperature, ten milliamps per square centimetre, glass H-cells, thin films on platinum-coated silicon. Background: commercial proton-exchange membrane electrolyzers run at roughly sixty to eighty degrees Celsius and at current densities a hundred times or more higher, inside a membrane-electrode assembly, and degradation can look quite different there.

Host 1And the thousand hours?

Host 2The paper's own introduction says industry needs thousands of hours, and a thousand is the start of that range, on a single long-term cell per material. The triplicate checks cover only the first two hundred and sixty-four hours.

Host 1What about the benchmark choice?

Host 2They compared against ruthenium oxide and plain palladium oxide made on their platform, not against iridium oxide, the durable incumbent. And their ruthenium oxide film delaminated, so it is not a strong reference for long-run stability.

Host 1Overpotentials, are they comparable to other papers?

Host 2Not directly. Screening values are not resistance-corrected, active surface area was not measured, and the authors say plainly the screen was designed for internal comparison across thousands of samples, not absolute numbers.

Host 1The additive story.

Host 2The additive is under a tenth of a percent, and the doped film already started with smaller crystals than the plain palladium oxide. The authors themselves call for controls that separate the effect of the additive from starting crystallinity, so the cause of the durability is not established.

Host 1Measurement limits?

Host 2X-ray fluorescence struggled to quantify indium and manganese in that much palladium, so screening-stage values for those elements are treated as comparative, roughly upper limits. Metal loss by X-ray fluorescence also has a detection floor, around zero point zero two micromoles per square centimetre.

Host 1The AI benchmarks.

Host 2Both are retrospective. The replay benchmark draws from a pool enriched by the agent's own live campaign, which the authors note favours it, and the palladium-discovery test treats every palladium system as equally good and uses a model trained on campaign data as its oracle.

Host 1And the language model baseline is a single model.

Host 2One model, off the shelf, in a simplified task, in a field where how you prompt matters. The result fits earlier reports, but it is a statement about that setup, not about language models in general.

Host 1Conflicts of interest?

Host 2All authors are employees who may hold equity, the company funded everything, and patents may be filed. That is normal for industrial research and openly disclosed, but it is also why independent replication matters. On the plus side, code and data for the main figures are posted on GitHub.

What to watch next

Host 1What should we watch for?

Host 2First, the controls the authors themselves list: amorphous palladium oxide, and crystalline palladium oxide with nickel or tantalum, to pin down what the additive really does. Second, tests in a real membrane-electrode assembly at operating temperature and much higher current.

Host 1And on the durability side?

Host 2Replicate runs to and beyond a thousand hours, an iridium oxide benchmark in the same setup, and measurements of dissolved palladium in the electrolyte during operation. If the needle structure is the key, someone should try to grow it on purpose, which the authors suggest is plausible by electrodeposition.

Host 1And for the AI side?

Host 2Live, prospective head-to-head campaigns rather than replays, which the authors admit would be costly. And, eventually, objectives beyond activity and stability, such as cost, supply resilience and recyclability, which they name as the next step.

Sign-off

Host 1That is the deep dive: AI-guided high-throughput discovery of iridium- and ruthenium-free palladium-oxide catalysts for durable acidic oxygen evolution, by Jenewein, Habib Zadeh and colleagues at Lila Sciences. The link is in the show notes.

Host 2Short version: a mostly automated lab screened nearly three thousand oxides and kept finding palladium oxide with dilute additives, and the best stayed under half a volt of overpotential for a thousand hours, against about two hundred for plain palladium oxide. It is a preprint, one long-term cell per material, room temperature and low current, with no iridium benchmark, and palladium is still a precious metal.

Host 1Every Friday this same feed carries the roundup episode: five new papers as trailers, and you pick which become deep dives. This has been the Hastrup Review of Science.