Skip to content
Advancing Bits to Master Atoms
← Every post

Guides8 min read

Planning a discovery program: budget, rounds and what to test first.

For R&D leads planning a quarter: start with a pilot, run in rounds, budget for the lab and spend compute where it changes a decision.

A long graphite rail carrying an even row of small identical silver crystals into the distance, one thin cobalt-blue disc beside it.A long graphite rail carrying an even row of small identical silver crystals into the distance, one thin cobalt-blue disc beside it.

A discovery program is a sequence of decisions under a budget: which target first, how many candidates to search, how many to make, when to check more deeply and when to hand off to someone who can make the material at scale. This guide is for the R&D lead planning a quarter of that work around MS Lattice. It sticks to what the record states about cost and time, and leaves lab rates and headcount to you.

Start with a pilot on your own target

The first step is a pilot. A pilot runs MS Lattice on your own target and hands you the shortlist, every candidate ranked, every check on record. Starting on your own target matters. A demonstration on someone else’s problem tells you little about how the engine handles your limits, your trade-offs and the language of your domain.

A pilot is scoped by a plan, and the quote follows it: the material domains, the campaigns and how your team will run MS Lattice. Before it starts, you agree with MatterSpace what is shared, where it is processed and how long it is kept. Bring a target where your team knows what a good answer looks like, so you can judge the shortlist on its merits.

Choose the mode by what you already have. With a target and no starting structure, Open Discovery searches the whole composition space. With a good material to improve, Prototype-Guided Optimization searches around your structure and ranks every variant against it. To watch the engine reach a material you know first, use Guided Rediscovery or the Blind Rediscovery Benchmark.

Think in rounds, not one big search

Discovery works in rounds. Each campaign returns a ranked shortlist; your lab makes and tests a few; what you learn tightens the next brief. MS Lattice is built for that rhythm: keep the winner, tighten the target, run again. The search learns where to look while it runs, so each round goes where the last one’s proof points.

A campaign can be short or long. On the record, an 8-candidate campaign driven by an AI agent ran in 4 to 5 minutes, a Guided Rediscovery campaign took about three minutes, and the blind benchmark searched 64 candidates in about 32 minutes. Campaigns can also run for weeks, with live progress round by round, checkpoints that let a long campaign pick up exactly where it stopped, and the option to stop at any time.

So plan the quarter around lab cycles, not compute cycles. Synthesis, characterization and the review that follows set the pace of a round.

What the compute costs, from the record

Compute is where AI discovery can quietly get expensive, because screening everything means paying for every structure that never had a chance. MS Lattice holds your requirements inside the search, so compute goes to contenders, never to a haystack. The record puts numbers on it.

  • ≈ $0.003of compute per candidate in the blind benchmark
  • ≈ 30 sof compute per candidate in the same run
  • < $0.20a blind rediscovery run: 64 candidates, 32 minutes, both sealed targets found
  • < $0.0004per candidate in the agent’s end-to-end campaign

At about a third of a cent of compute per candidate, the arithmetic at scale is simple: 1,000 candidates for $3, 10,000 for $30, 100,000 for $300 and 1,000,000 for $3,000. An AI agent’s complete 8-candidate campaign came in under $0.003. Ranking the right structure from a formula took about 38 seconds per compound.

Every plan shows its estimated cost before launch, and a plan that cannot run is refused with the reason before anything starts. Every campaign runs inside the budget you set, and an agent driving MS Lattice over MCP checks the budget and the price before it starts. Those figures are the compute on the record; what a pilot through Norck costs follows the pilot plan and comes as a quote.

One small precision-milled cobalt-blue disc resting on a matte slate plane beside a tall ordered crystal tower of matte graphite and satin silver spheres.One small precision-milled cobalt-blue disc resting on a matte slate plane beside a tall ordered crystal tower of matte graphite and satin silver spheres.
A tiny crystal on a precision balance: the compute behind one candidate weighs little next to the work of making it.

Where the real cost sits

Set the compute figures against what follows them. Making a new inorganic material means sourcing precursors, finding a route that forms the right phase, confirming the structure, then measuring the property that matters under conditions close to service. A coating has to be deposited and cut with. A magnet has to be made, processed and measured at temperature. An electrolyte has to be built into a cell. Each step takes skilled people, instruments and time, and each costs far more than the search that proposed the candidate.

That is the case for a short, well-ranked list. A list of millions moves the problem from the computer to the bench. The planning consequence is direct: budget the program around synthesis and testing, and treat compute as the small line that decides where the large one goes.

Your lab can make hundreds. MS Lattice delivers the ones worth making.

How many to make per round

The right number per round is the number your lab can make and characterize well in one cycle. A candidate made carelessly, in the wrong phase or at poor density, tells you about the process rather than the composition, and wastes a slot.

  • Size the round to capacity: count the synthesis routes, furnace or deposition time and characterization slots your team has, then fill them.
  • Spread the bets: take candidates from different families near the top, since MS Lattice aims for different strong answers, not near-copies.
  • Weight by margin: favor candidates whose advantage on your key property is clear of its uncertainty.
  • Include a baseline: make your current material or a known reference alongside, so every measurement has a comparison your team trusts.
  • Hold some back: keep the next few ranked candidates ready, so a failure in the first batch does not stall the round.

Early rounds can lean wider, to learn where the estimates and your process agree. Later rounds narrow as the targets tighten and Prototype-Guided Optimization takes over from your best made material.

When to call first-principles verification

First-principles verification is the check you call when a verdict matters: stability against every rival phase of the chemistry, elastic constants and phonons, on demand, for the candidates you choose. On the record, it called stability right four times out of four and reproduced silicon on five reference values out of five, each within the tolerance set before the run.

In a plan, it sits at the boundary between screening and the lab. Call it for candidates about to take serious synthesis effort, when stability decides between two close candidates, and when your application depends on how a material carries load or how its lattice vibrates. What does not hold is caught before a single day in the lab.

When to bring in Norck

Some teams have the lab to take a shortlist forward. Others do not, or want to go further than a lab sample. Norck runs MS Lattice under agreement with MatterSpace, then synthesizes, validates and manufactures the winner. A material is not discovered until it is made.

Norck works two ways, existing material optimization to make what you have better and breakthrough material discovery to find what does not exist yet, and its work is offered in scopes, so you can buy only the part you need:

  • Discovery: new material candidates, or an optimized existing material.
  • Discovery + Validation: candidates synthesized and physically tested.
  • Discovery + Validation + Production: a validated material, produced.
  • End to end: from discovery to your final, manufactured parts.

Real-world results feed straight back into the next round of design. In a plan, bring Norck in when lab capacity is the constraint, when a round needs synthesis and physical testing you do not run in-house, or when a winner is ready to move toward production and parts.

Eight circular matte graphite and satin silver stepping plinths rising in a gentle curve from a cobalt-blue dot on the left to a precision-machined satin metal part on the right.Eight circular matte graphite and satin silver stepping plinths rising in a gentle curve from a cobalt-blue dot on the left to a precision-machined satin metal part on the right.
Powder, pressed pellet, rough blank and machined part: the path a winning composition still has to travel after the shortlist.

A sample plan, in outline

Every program differs, so this is a shape, not a schedule. It uses only what the record states and leaves durations and lab costs to your own numbers.

  1. Pilot on your own targetAgree the plan and how your data is handled, then run MS Lattice on a target your team understands. Read the shortlist, the scores and the record against what you already know.
  2. Run the first round of campaignsRun Open Discovery, or Prototype-Guided Optimization from your current material, with requirements and preferences set from what the pilot taught you. Check the price before launch and set the budget.
  3. Verify, then makeSend the candidates headed for the lab through first-principles verification. Make a round sized to your lab, spread across families, with a baseline beside it.
  4. Feed back and tightenBring the measurements and the reason for each miss back into the brief. Keep the winner, tighten the target, run again.
  5. Decide the hand-offWhen a candidate holds up in your tests, decide whether your team carries it forward or Norck takes it to validation, production or finished parts.

The compute in that plan stays small, and you see its price before each launch. The decisions that cost real money, what to make and how far to take it, are what the record of every check is there to support.

Questions teams ask

What should we budget for compute?

On the record, about a third of a cent per candidate, and every campaign runs inside the budget you set.

How long does a round take?

Campaigns run for minutes or for weeks. Synthesis and testing set the pace of a round.

How is a pilot priced?

The quote follows the pilot plan: the material domains, the campaigns and how your team will run MS Lattice. Write with your target, and MatterSpace comes back with a plan and a quote.

Can we start from the material we use today?

Yes. Prototype-Guided Optimization searches around your structure and ranks every variant against the original.

Can we buy only part of Norck’s process?

Yes. Discovery, validation, production or end to end.

MatterSpace’s first domain engine

MatterSpace built MS Lattice for the scientists and engineers who discover and design new materials, starting from the properties their product needs.

Keep reading

Have a materials target? Tell us what it must do.

Share the material domain, the properties you need and the limits the material must respect. We will show you how MS Lattice would approach it through Norck, from a ranked shortlist for your own lab all the way to finished parts.