Product 2 of 5 · Design + active learning

2InventaDesign + active learning

Pick the next experiment by what it will teach you

Inventa turns a property target into candidate structures, then chooses which ones to simulate or make next by balancing predicted value against uncertainty. It spends GPU-hours and lab weeks where they change the answer.

Abstract

Role
Design + active learning
Loop step
2 · choose candidates
v1 proposals
Enumeration + substitution
Later
Generative inverse design
Example · Inventa campaignExample
Targetσ > 1 mS/cm · Ehull ≤ 25 meV/atom · no Co
Space48,000 substitutions on Li₆PS₅Cl
Round2 of 4 · 212 above target
Next batch8 exploit + 4 explore
Budget used286 of 400 GPU-h
Example. Illustrative campaign state. Explore picks are high-uncertainty candidates chosen to teach the model, not because they are predicted best.

01 · Capabilities

What Inventa does

Target specs

Properties with units

Targets and limits written once and tracked through every round.

Substitution search

From known prototypes

Enumerates allowed swaps around structures that are known to form.

Active learning

Exploit and explore

Bayesian selection that mixes likely winners with candidates the model is unsure about.

Budget planner

GPU-hours and lab slots

Plans a round within a compute budget and the batch size your lab can take.

Synthesizability gate

Before the lab

Checks hull distance and known precursors before a candidate is proposed for synthesis.

Generative design

Target to structure

Proposes new structures from a property target. Planned

02 · Method

How Inventa works

  1. 1

    Set the target

    Property, limit, units and exclusions.

    Input Target spec
  2. 2

    Propose candidates

    Enumerate substitutions; size the space and the budget.

    Time Hours
  3. 3

    Score with the loop

    Send to Equiva and Dynara; collect predictions and uncertainties.

    Uses Equiva, Dynara
  4. 4

    Select a batch

    Rank by expected improvement; propose for approval.

    Gate Named approver

03 · Catalogue

Filter the candidate list

Class
Status

10 of 10 candidates shown

Example candidates across three programs. σ in mS/cm, Ehull in meV/atom, u = model uncertainty in eV/atom.
IDClassStatusσEhulluDecision
cand-0118ElectrolyteStable12.480.06Exploit
cand-0342ElectrolyteStable4.1140.21Explore
cand-0731ElectrolyteStable0.8110.09Hold
cand-1207CathodeStable—60.04Exploit
cand-1288CathodeOff hull—610.12Drop
cand-2051CatalystStable—190.30Explore
cand-2093CatalystOff hull—880.08Drop
cand-0409ElectrolyteOff hull6.3420.17Drop
cand-1544CathodeStable—30.05Exploit
cand-2210CatalystStable—240.11Hold

04 · Specification

Specification

First-release scope. Items marked planned are not in the first release.
ItemDetail
InputsTarget spec, structure prototypes, allowed substitutions, budgets
OutputsRanked candidates, proposed batches with reasons, round-by-round target tracking
SelectionExpected improvement and upper-confidence-bound strategies; batch diversity constraint
GatesHull threshold, element exclusions, precursor availability, human approval
Generative proposalsDiffusion or flow models over crystal structures Planned
ProvenanceEvery proposal linked to the model version and data that produced it

05 · Place in the loop

Where Inventa sits

Inventa is the decision step. It takes the question Mentra framed, asks Equiva and Dynara for predictions, and decides what Synthex should make. It is also where the loop learns fastest: one well-chosen uncertain candidate can correct a model more than ten safe ones.

1MentraResearch copilot

Reads the literature and your notebooks, frames the question and drafts the campaign.

2InventaDesign + active learning

Turns the target into candidates and picks the next batch by expected information.

3EquivaInteratomic potentials

Predicts energies, forces and stresses at near-DFT accuracy, with an uncertainty for each.

4DynaraGPU simulation

Relaxes, scores against the hull and runs GPU molecular dynamics on the survivors.

5SynthexSelf-driving lab

Sends the short list to partner or in-house labs and brings measurements back.

Return pathSynthex measurements → fine-tune Equiva → better predictions in the next round.

06 · Questions

Questions about Inventa

Is generative design available?

Not in the first release. We start with enumeration and substitution around known prototypes, which already covers large spaces, and add generative proposals once they beat that baseline on our benchmarks.

Why would we test a candidate that is not predicted best?

Because the model is unsure about it. Measuring it reduces uncertainty for many similar candidates at once, which makes the next round's ranking more reliable.

Can we override a selection?

Always. You approve every batch and can add or remove candidates; Inventa records the change and its reason.

Next step

Test Inventa on one of your chemistries

The 2027 pilot pairs one of your programs with our scientists for twelve weeks. You bring a target and reference data; we bring the GPUs and the models.