Notebook · materials discovery on GPUs

Simulate the material before you make it

Forgite predicts energies and forces at near-DFT accuracy on GPUs, screens tens of thousands of candidates against your target, and sends only the short list to the lab.

Status
Pre-launch
Pilot
Three design partners, 2027
First verticals
Batteries, catalysts
Runs on
NVIDIA GPUs, cloud or on-prem

Question

Which of 48,000 candidate solid electrolytes are stable and conduct lithium fast enough to be worth making?

Method

Relax each one with an Equiva potential, measure its distance to the convex hull, then run GPU molecular dynamics on the survivors.

Result

A ranked short list for the lab, each candidate with its uncertainty, in days rather than quarters. (design target)

Fig. 1 · Solid-electrolyte screen, 240 candidatesExample
Scatter plot of 240 example candidates: energy above hull against ionic conductivity. Candidates left of the hull threshold and above the conductivity target are kept.02550751001251500.0010.010.1110100Energy above hull, meV/atomIonic conductivity, mS/cm (log)
133stable within threshold
107off hull, screened out
28kept for MD and the lab queue
Top of the list: cand-0026Ehull 4 meV/atom · σ 54.8 mS/cm
  • Kept: stable and above target
  • Stable, below target
  • Off hull
  • Highest conductivity kept
Example. Synthetic data drawn to show the method, not results from a real campaign. Move the two limits to see how the hull threshold and the property target decide what goes forward.

01 · Design targets

What the platform is built to do

Design target~1,000×

throughput of an ML potential relaxation compared with a DFT relaxation of the same cell

Design target10⁵candidates

screened per campaign on one GPU node in a working week

Design target≤30meV/atom

mean absolute error on held-out structures in a fine-tuned chemistry

Example1 in 200

share of a typical screen that might reach the lab queue

02 · Problem

Why new materials still take decades

Density functional theory is accurate enough to trust and too slow to use at scale: a single relaxation of a few hundred atoms can take hours of CPU time, so a team can afford to compute a few hundred candidates per quarter. The rest of chemical space goes unexamined.

Machine-learned interatomic potentials close most of that gap, but they are mostly research code. They are hard to fine-tune, rarely report their own uncertainty, and are not connected to the decision of what to make next. Meanwhile each synthesis and characterisation round in the lab takes weeks.

The result is a cycle measured in years per material, at the exact point where batteries, catalysts and chips need new ones.

03 · The loop

Five products, one discovery loop

Each product does one job. They share the same potentials, the same data store and the same provenance graph, so every round of the loop improves the next.

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.

04 · Method

How one campaign runs

  1. 1

    State the target

    A property with units and a limit: ionic conductivity above 1 mS/cm at 25 °C, stable within 25 meV/atom of the hull, no cobalt.

    Owner Your program leadTime One working session
  2. 2

    Generate candidates

    Inventa enumerates substitutions and, later, generative proposals around known structure prototypes.

    Owner InventaTime Hours
  3. 3

    Predict and screen

    Equiva relaxes every candidate; Dynara scores stability against the hull and runs MD on the survivors.

    Owner Equiva + DynaraTime Days on GPUs
  4. 4

    Choose what to make

    Active learning ranks by expected value and uncertainty; you approve the batch for synthesis.

    Owner Inventa + your teamTime One review
  5. 5

    Measure and learn

    Synthex returns XRD, impedance or activity data, which fine-tunes the potential for the next round.

    Owner Synthex + lab partnerTime Weeks per batch

05 · Who it is for

Built for teams that already run DFT

Cathodes, electrolytes, anodes

Battery developers

Screen solid electrolytes and cathode substitutions for conductivity and voltage before committing furnace time.

Hydrogen, CO₂ conversion

Catalyst programs

Rank surfaces and dopants by adsorption energies and stability under reaction conditions.

Dielectrics, barriers, precursors

Semiconductor materials

Narrow candidate films and precursors by band gap, diffusion and thermal stability.

06 · Stack

Runs on the NVIDIA scientific stack

PotentialscuEquivarianceEquivariant tensor products for training and inference.
ScreeningALCHEMIMaterials property prediction and batched relaxation.
SimulationWarp + CUDACustom MD and property kernels.
ServingTriton + TensorRTBatched inference over millions of structures.
CopilotNeMo + NIMRetrieval over literature and your data, with guardrails.
DataRAPIDSGPU dataframes for campaign results.
Lab edgeJetson OrinOn-instrument triage of characterisation data.
HardwareDGX / HGXTraining foundation potentials; MD ensembles.

07 · For developers

Start a screen in a few lines

python · forgite SDK (planned interface)
from forgite import Campaign

campaign = Campaign.create(
    name="sse-li-argyrodite",
    space="substitutions:Li6PS5Cl",           # prototype + allowed swaps
    potential="equiva/battery-ft",
    filters={"e_hull_mev": 25, "exclude": ["Co"]},
    target={"ionic_conductivity_mS_cm": ">1"},
)
campaign.run(gpus=8)
for c in campaign.top(20):
    print(c.id, c.e_hull, c.sigma, c.uncertainty)

08 · Questions

Questions R&D teams ask

How accurate are the potentials?

Our target is a mean absolute error at or below 30 meV/atom on held-out structures in a fine-tuned chemistry, reported with a per-prediction uncertainty. Accuracy outside the training distribution is lower, which is why every prediction carries its uncertainty and why uncertain candidates are flagged rather than ranked.

Do we have to share our data?

No. Fine-tuning runs per tenant, and your structures, DFT results and lab data stay in your store. Shared foundation models improve only from data you explicitly opt in to share.

Do you replace DFT?

No. Forgite does the wide screen so that DFT and the lab are spent on the few candidates that deserve them. Spot-check DFT on the short list is part of the default method.

Is this available today?

Not yet. Forgite is pre-launch. We are recruiting three design partners for a twelve-week pilot in 2027.

Next step

Bring one target and your reference data

We will run the first screen with you, show every assumption, and leave you with a ranked list and the notebook that produced it.