How it works

A campaign is an experiment you can rerun

Every campaign is written down like a notebook entry: the question, the search space, the model version, the filters and the result. Anyone on your team can rerun it and get the same list.

Abstract

Steps
6
Typical length
2–6 weeks per round
Human gates
2
Output
Ranked list + notebook

01 · Procedure

Six steps, two human decisions

  1. 1

    Frame the question

    Mentra helps write the target with units and limits and pulls prior results from literature and your archive.

    Owner Program leadTime Day 1
  2. 2

    Define the search space

    Pick structure prototypes and allowed substitutions. Inventa estimates the size of the space and the GPU budget.

    Owner Computational chemistTime Day 1–2
  3. 3

    Relax and score

    Equiva relaxes each candidate; Dynara computes the distance to the convex hull and discards anything above your threshold.

    Owner Equiva + DynaraTime Days 2–5
  4. 4

    Simulate the survivors

    GPU molecular dynamics and property calculators estimate conductivity, voltage or adsorption energies, with confidence intervals.

    Owner DynaraTime Days 5–10
  5. 5

    Approve a batch

    Inventa proposes a batch that balances the best predictions with the most informative uncertain ones. A named person approves it.

    Owner Your teamTime One review
  6. 6

    Measure and fine-tune

    Synthex sends the batch to the lab and returns measurements, which fine-tune the potential before the next round.

    Owner Synthex + labTime Weeks 3–6

02 · Stability

The hull decides what is worth simulating

Fig. 2 · The convex hull decides stabilityIllustrative
Convex hull of formation energy for a binary system, with five candidates above it and their distance to the hull in meV/atom.0.0-0.2-0.4-0.60.000.250.500.751.001844956631Fraction of B in A₁₋ₓBₓFormation energy, eV/atom
  • On the hull (stable)
  • Within 25 meV/atom, kept
  • Further above, screened out
Illustrative. A made-up binary system. Each number is the vertical distance to the hull in meV/atom; a candidate on the hull cannot lower its energy by splitting into neighbours.

03 · Responsibilities

Who does what

TaskForgiteYour teamLab partner
Target and constraintsDrafts with MentraOwns and approves—
Search spaceEnumerates and sizesChooses prototypes—
Prediction and screeningRuns on GPUsReviews outliers—
Batch for synthesisProposes and explainsApprovesConfirms feasibility
Synthesis and measurementRoutes jobs, captures dataSets protocolsPerforms
Model updateFine-tunes per tenantApproves the new version—

04 · Configuration

A campaign, written down

campaign.yaml · example
name: sse-li-argyrodite-r1
question: Which Li6PS5Cl substitutions keep sigma above 1 mS/cm?
space:
  prototype: Li6PS5Cl
  substitute: { S: [Se, O], Cl: [Br, I], P: [Si, Ge, Sb] }
potential: equiva/battery-ft@0.3.1        # pinned for reproducibility
filters:
  e_hull_mev_per_atom: 25
  exclude_elements: [Co, Ge]   # cost and supply rule
simulate:
  md: { temperature_K: 600, time_ps: 1000, seeds: 3 }
select:
  batch_size: 12
  strategy: expected-improvement
budget_gpu_hours: 400

05 · Outcome

What a round leaves behind

EnumeratedSubstitutions on three sites
48,000
RelaxedConverged with Equiva
46,900
StableWithin 25 meV/atom of the hull
9,400
SimulatedMD at 600 K, three seeds
1,100
Above targetσ > 1 mS/cm, low uncertainty
212
Approved batchSent for synthesis
12

Example. Counts are illustrative for one round of a solid-electrolyte campaign. Bars use a log scale so the last stages stay visible.

06 · Questions

Questions R&D teams ask

What if the potential is wrong for our chemistry?

Uncertainty rises before accuracy falls. Inventa routes high-uncertainty candidates to DFT spot checks or to the lab on purpose, and the results fine-tune the potential. You see the error on your own held-out set before you trust a ranking.

How long does a round take?

Screening and simulation take days on GPUs. The lab sets the pace after that, usually two to six weeks per batch.

Can we stop after the screen?

Yes. Many teams will use the ranked list and run synthesis their own way. Synthex is optional.

Next step

Run one round with us

The pilot covers one full round: question, screen, simulation, a batch for your lab, and a model update from the results.