How it works
A search problem where every read is expensive
Written to be understood without a chemistry background. If you think in systems terms, there are analogies throughout — they are approximate, but they are the right shape.
The problem in one line
Catalyst researchers search a combinatorial space where every measurement costs months — and the metric they optimise on is not the metric that decides the outcome.
What the search actually is
A PEM electrolyser splits water into hydrogen using a catalyst. That catalyst has to do three things at once, and they pull against each other.
1 · Low overpotential
How much extra voltage is wasted driving the reaction. It converts directly into electricity cost — about 26.6 kWh per kilogram of hydrogen for every volt.
2 · Survives acid at voltage
Acidic oxygen evolution is destructive. Plenty of highly active materials dissolve in weeks, which is a cost problem as much as a technical one — degradation forces earlier stack replacement.
3 · Little iridium
World production is roughly 7–8 tonnes a year, as a by-product of platinum mining. Supply responds to platinum demand, not to price. At gigawatt scale this is a physical ceiling, not a line item.
Easily millions of candidates. A laboratory can physically synthesise and test perhaps five to twenty a year.
The pipeline today — and what we replace
| Step | How it is done now | What it costs |
|---|---|---|
| Shortlist candidates | A scientist reads the literature and uses judgement | 2–3 weeks, ~5 candidates, little record of why |
| Check computationally | Density functional theory — needs an HPC cluster, a computational chemist, and licences that run into five figures | Days to weeks per candidate. Most small groups skip it entirely |
| Synthesise and test | Laboratory work | 2–3 months per batch |
| Decide if it is commercially worth anything | Informally, in a spreadsheet, often by a different team | Late, inconsistent, and frequently never done properly |
Because the computational step is expensive, it gets skipped — so groups go straight from intuition to the laboratory, spending their scarcest resource on candidates a cheap filter would have rejected. And because the economic step comes last, the whole search optimises activity, which is easy to measure, instead of cost per kilogram, which is what decides whether anyone buys the result.
The translation, if you think in systems terms
DFT is a full table scan
Exact results, and far too slow to run across the space. You would not query a billion rows this way, and nobody screens a composition space this way either.
An MLIP is an approximate index
A neural network trained on DFT outputs, predicting the same energies roughly 10,000× faster. Wrong in the third decimal, right about the ordering — and ordering is the whole job when you are ranking.
LCOH is a computed column
Physical properties → cell voltage → kWh/kg → iridium g/kW → euros. A materialised view over the technical metrics, expressed in the unit the decision is actually made in.
So the product is: an index over a space that previously only supported full scans, plus a computed column nobody had.
There is a third piece: lineage. Every result carries its model version, method parameters and input and output hashes, and can be re-run pinned to the version that produced it. Reproducibility as a first-class column rather than an afterthought — which matters when a result has to survive a reviewer, a patent examiner, or a colleague six months later.
The flow, before and after
The question: can we cut iridium by half without losing performance?
Before
read literature 3 weeks pick 5 candidates by intuition synthesise + test 3 months -> 4 fail, 1 marginal cost assessed later, elsewhere
One iteration: ~4 months, 5 samples, and no record of the candidates you did not pick.
After
define a space (40-200) minutes
screen all of them minutes-hours
-> activity, stability,
iridium intensity, LCOH
-> ranked on a cost frontier
-> out-of-domain refused
pick top 3-5, reasoning attached
synthesise + test 3 months
feed results back Same laboratory time — now spent on candidates that survived a filter, ranked on the commercial objective from the start.
What we are explicitly not replacing
The laboratory
Experiments still decide. This is triage, and triage is not a verdict.
DFT for final numbers
You still want converged calculations for a paper or a specification. We replace DFT for the step where you are deciding what deserves DFT.
Domain expertise
It narrows the field. A chemist still judges what comes out of it, and should.
Where you should push back
The fair objections, stated plainly rather than buried.
General-purpose interatomic potentials are trained predominantly on bulk crystals, and published benchmarks show roughly double the error on oxide-rich surfaces with oxygen adsorbates compared with the metallic majority of the training data. Acidic oxygen evolution is precisely that hard case. It is the reason this platform specialises rather than covering all materials — and the gap is a thing to prove with a number, not to assert.
It exists to compare options on a consistent basis. It is not a bankable financial model, and the ranges it reports are sensitivity bands rather than confidence intervals. The full method is published and versioned so you can check it.
Declining to answer is better than returning a plausible wrong number, but it does not make the underlying model more accurate. It makes its limits visible.
Until a laboratory confirms a candidate this approach surfaced, it is a well-argued hypothesis. We would rather say that than imply otherwise.
Questions
What problem does computational catalyst screening solve?
Catalyst discovery searches a space of millions of candidate compositions, structures and surfaces, but a lab can physically synthesise and test perhaps five to twenty a year. Screening computationally narrows that space cheaply, so the scarce laboratory time is spent on candidates that already survived a filter rather than on intuition alone.
What is a machine-learned interatomic potential?
A neural network trained on the outputs of density functional theory. It predicts the same energies roughly ten thousand times faster, approximately rather than exactly. It is wrong in the third decimal and right about the ordering, which is what matters when the job is ranking candidates rather than producing a final number.
Does this replace DFT?
No. It replaces DFT for the step where you are deciding which candidates deserve DFT. You still want converged density functional theory for a final number in a paper or a specification. Screening is triage; confirmation is a different job.
Why does cost per kilogram matter more than activity?
Because activity is only one of three things that decide whether a catalyst is commercially viable, alongside stability and iridium intensity. A hundred millivolts of overpotential is worth roughly twelve cents per kilogram of hydrogen at typical electricity prices. Using a third less iridium can decide whether a design can be built at gigawatt scale at all. Ranking on activity alone optimises the variable that is easiest to measure rather than the one that decides the outcome.
Why is iridium the constraint?
World primary production is roughly seven to eight tonnes a year, and it is a by-product of platinum mining. Because it is a by-product, supply responds to platinum demand rather than to iridium demand or price, so it cannot scale quickly. At gigawatt deployment scale that becomes a physical ceiling rather than a cost line.
How accurate is screening, honestly?
Accurate enough to rank and triage, not accurate enough to be a final number. General-purpose interatomic potentials are measurably least accurate on oxide surfaces with oxygen adsorbates, which is exactly the acidic oxygen-evolution regime. That gap is the reason this platform specialises rather than covering all materials, and where a candidate falls outside the validated domain the system refuses to answer rather than guessing.
See the arithmetic yourself
The LCOH calculator turns catalyst performance into cost per kilogram, free and with no account. It is the smallest honest version of the whole argument above.