Domain packs
One engine. Focused per problem.
A screening tool that is vague about the science is useless, and a tool that only ever works for one chemistry is a dead end. So the engine is general and every pack is narrow — deliberately.
What every pack has in common
These are the same regardless of field, and they are what the platform is.
A space too big to test
Candidates run to thousands or millions. A laboratory gets through a handful a year.
Experiments that cost months
Whether it is a synthesis batch or a growing season, you cannot brute-force it.
Models that need infrastructure
The methods exist and are often open. Clusters, licences and specialists are what most labs lack.
A deciding metric computed too late
Cost, yield, viability — assessed after the work, so the search optimises the easy variable instead.
A pack supplies what is specific: how a candidate is represented, which models predict its properties, where those models stop being valid, and the economic model that turns a predicted property into a decision. Everything else — the queue, the cache, the provenance chain, the refusal behaviour, the ranking and the frontier — is shared.
The packs
One is live. The rest are honest about their status.
PEM electrolysis catalysts
Acidic oxygen-evolution catalysts for green hydrogen — activity, stability and iridium intensity, closed to cost per kilogram.
- Space
- Compositions, structures and surfaces of acidic oxide catalysts
- Experiment
- Synthesis and electrochemical testing — 2–3 months per batch
- Predictor
- Fine-tuned machine-learned interatomic potentials
- Decides on
- Levelised cost of hydrogen (€/kg)
Explore this pack →
Ammonia cracking
Catalysts for releasing hydrogen from ammonia — the carrier problem, where hydrogen is shipped as ammonia and cracked on arrival.
- Space
- Supported metal catalysts and operating conditions
- Experiment
- Catalyst synthesis and reactor testing
- Predictor
- Interatomic potentials plus reaction-network models
- Decides on
- Cost per kilogram of delivered hydrogen
See what it would do →
Methane pyrolysis
Splitting methane into hydrogen and solid carbon, avoiding CO2 at the point of production.
- Space
- Catalyst and molten-media systems
- Experiment
- Reactor campaigns
- Predictor
- Interatomic potentials and thermodynamic models
- Decides on
- Cost per kilogram, plus carbon co-product value
See what it would do →
CO2 conversion and e-fuels
Turning captured CO2 into fuels and chemicals — selectivity is the hard constraint, not just activity.
- Space
- Electrocatalysts and thermocatalysts, with product selectivity
- Experiment
- Electrochemical and reactor testing
- Predictor
- Interatomic potentials with selectivity descriptors
- Decides on
- Cost per tonne of product
See what it would do →
Why the engine is not hydrogen-specific
Every pack above is a molecular-energy problem, and that is deliberate — it is the field we know, the audience we serve, and depth is what makes a screening result worth acting on.
But nothing in the engine is. Define a candidate space, predict, rank on what actually decides, refuse where the models are out of their depth, record provenance, feed results back — that loop holds anywhere experiments are slow and candidates are many. The pack supplies the science; the engine supplies the discipline.
Built molecule-agnostic, focused on molecular energy. The first is an architecture decision, the second is a commercial one, and they are allowed to differ.
Which pack comes next is decided by who asks. If your laboratory has more candidates than capacity to test them, that is the conversation worth having.
Questions
What is a domain pack?
The platform underneath is the same everywhere: define a candidate space, predict, rank on the objective that actually decides, refuse where the models are out of their depth, record provenance, feed results back. A domain pack supplies what is specific to one field — the candidate representation, the prediction models, the validity limits, and the economic model that turns a predicted property into a decision.
Why not build one general tool for all of science?
Because general-purpose models are weakest exactly where a field is hardest, and a general tool has no incentive to fix that. Depth is what makes a screening result trustworthy enough to spend laboratory time on. The engine generalises; the science must not be watered down.
Is the platform limited to hydrogen?
The packs today are all molecular-energy problems, and that focus is deliberate. The engine underneath is not chemistry-specific — the loop of defining a candidate space, predicting, ranking on what decides, refusing where the models are out of their depth and recording provenance holds wherever experiments are slow and candidates are many. We build depth where we have it rather than breadth we cannot defend.
Can you build a pack for my field?
Possibly. The honest requirement is a usable prediction model for the field, a candidate space that is genuinely too large to test, experiments expensive enough that triage is worth paying for, and a decision metric we can compute. If those hold, tell us — early packs are chosen by who asks.
Your field isn't on the list?
Early packs get chosen by who asks. If your laboratory has more candidates than capacity and a prediction model that mostly works, we would like to hear what you would screen first.
Tell us what you would run