Product 3 of 5 · Interatomic potentials

3EquivaInteratomic potentials

Interatomic potentials that report their own uncertainty

Equiva is the model layer: equivariant graph neural network potentials that predict energies, forces and stresses at near-DFT accuracy, fine-tuned on your chemistry and served on GPUs.

Abstract

Role
Interatomic potentials
Loop step
3 · predict
Predicts
E, F, σ with uncertainty
Built on
cuEquivariance, DGX/HGX
Fig. 1 · Held-out parity, battery fine-tuneExample
Parity plot: predicted against reference formation energy for 90 example structures, clustered on the diagonal.-3-3-2-2-1-100DFT reference, eV/atomEquiva prediction, eV/atom
30meV/atom MAE, design target for the battery fine-tune
90held-out structures
±σreported per prediction
  • Held-out structure
  • Perfect agreement (diagonal)
Example. Synthetic points around a design-target error. Real accuracy reports are produced per chemistry against your own reference data.

01 · Capabilities

What Equiva does

Foundation potential

Broad coverage

Pretrained on public and licensed DFT data across the periodic table.

Vertical fine-tunes

Batteries first

Fine-tuned packs for battery materials, then catalysts and semiconductors.

Your fine-tune

Per tenant

Fine-tune on your DFT and experimental data; the result stays in your tenant.

Uncertainty

Ensemble estimates

A calibrated uncertainty with every energy and force, used to flag and route candidates.

Accuracy reports

Per chemistry

Error on held-out structures against DFT and, where available, experiment.

GPU serving

Batched endpoints

Low-latency endpoints for relaxation and MD, served with Triton and TensorRT.

02 · Method

How Equiva works

  1. 1

    Pretrain

    Train the foundation model on large DFT datasets on DGX/HGX.

    Uses cuEquivariance
  2. 2

    Fine-tune

    Adapt to a vertical or to your data, keeping a held-out set aside.

    Time Hours to days
  3. 3

    Calibrate

    Fit uncertainty so that predicted error matches observed error.

    Check Calibration plot
  4. 4

    Serve

    Publish a pinned version as a GPU endpoint used by Dynara and Inventa.

    Uses Triton, NIM

03 · Specification

Specification

First-release scope. Items marked planned are not in the first release.
ItemDetail
ArchitectureEquivariant message-passing graph neural network
PredictsTotal energy, atomic forces, stress tensor; uncertainty per prediction
Accuracy target≤ 30 meV/atom MAE on held-out structures in a fine-tuned chemistry (design target)
Training dataPublic DFT datasets, licensed data where agreed, tenant data only with consent
VersionsSemantic versions; campaigns pin a version
Synthetic augmentationGenerated structures to cold-start sparse chemistries Planned

04 · Place in the loop

Where Equiva sits

Equiva sits in the middle of the loop. Inventa asks it for predictions, Dynara runs dynamics with it, and Synthex's measurements come back to fine-tune it. It is the part of the platform that gets better with every round.

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.

05 · Questions

Questions about Equiva

How does this compare with DFT?

Equiva is trained to reproduce DFT, so it inherits DFT's systematic errors and adds its own. It is for screening at scale, with DFT spot checks on the short list.

What if our chemistry is unusual?

The uncertainty will tell you. High uncertainty routes candidates to DFT or the lab, and those results fine-tune the model for your space.

Can we export the model?

Fine-tuned weights trained on your data are yours. Export terms for on-premises use are part of the commercial agreement.

Next step

Test Equiva 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.