Orb Universal Interatomic Potential
A universal interatomic potential (graph network) for atomistic modelling of materials; denoising diffusion pretraining only initialises the potential, which predicts energy, forces and stress.
The loop
- Proposes
Orb is a universal interatomic potential, not a generative search; denoising pretraining only initialises the potential.
- Judged by
The potential predicts energy, forces and stress.
- Steers
No generative search; the model is trained in two phases, denoising diffusion pretraining and then the interatomic potential.
Headline result
Topped the Matbench Discovery leaderboard for predicting stable inorganic crystals at 3-6x higher relaxation speed than prior equivariant potentials.
What this does not prove
Learned universal potentials lack explicit long-range electrostatics unless augmented with Ewald summation terms.
How it generalizes the harness
- A domain authority judges
- Feedback retrains the generator
- Search over a population
- The generator is not an LLM
A universal interatomic potential, trained first as a denoising diffusion model and then as a neural network potential; it is not a generative search.
Classification
| What it proposes | Crystal structure |
|---|---|
| Generative model | Graph network interatomic potential, initialised from a denoising diffusion pretraining phase |
| Algorithmic authority | None in the loop: the potential predicts energy, forces and stress, and it is not a generative search |
| Feedback | New labeled data |
| What gets updated | Model weights and training data |
| When it runs | During training |
| Validity | Approximate physics (simulation, not lab results) |
| Evidence | Code released |
| Lineage | Original system |