MatterGen
A joint 3D diffusion model generates atom types, coordinates and periodic lattices conditioned on target properties, and Density Functional Theory relaxations verify stability and property values.
The loop
- Proposes
3D periodic crystal structures (atom species, fractional coordinates and lattice vectors) sampled by an equivariant diffusion model with property adapter modules
- Judged by
Machine-learned interatomic potential pre-screening followed by Density Functional Theory (DFT) structural relaxation, convex-hull stability and property calculation
- Steers
Adapters are fine-tuned on a labelled dataset; generated candidates are pre-relaxed with MatterSim and checked with DFT afterwards.
Headline result
In Nature (2025), MatterGen more than doubled the fraction of generated inorganic structures that are both novel and thermodynamically stable (within 0.1 eV/atom of the DFT convex hull) compared with prior generative baselines, and reduced root-mean-square atomic displacement during DFT relaxation by more than 50%. In a targeted high-bulk-modulus search, the authors synthesized a generated compound (TaCr2O6) in the lab whose measured bulk modulus (169 GPa) matched the target within 20%.
What this does not prove
Stability and property rates are primarily evaluated in 0 K DFT simulations rather than wet-lab synthesis; only one candidate structure (TaCr2O6, which turned out to be a disordered solid solution in experiment) was synthesized in the paper.
How it generalizes the harness
- A domain authority judges
- Feedback retrains the generator
- Search over a population
- The generator is not an LLM
- The output outlives the model
- Multi-fidelity screening ladder
An equivariant 3D diffusion model generates crystals directly in continuous and discrete symmetry spaces, and DFT physics calculations check generated candidates afterwards, while the adapters are fine-tuned on a labelled dataset.
Classification
| What it proposes | Crystal structure |
|---|---|
| Generative model | Equivariant score-based diffusion model over atom types, coordinates and periodic lattices with fine-tuned adapter modules (not an LLM) |
| Algorithmic authority | MatterSim pre-relaxation and Density Functional Theory (DFT) convex-hull and property evaluation, plus X-ray diffraction and nanoindentation on the synthesized sample |
| Feedback | Checks at several fidelities |
| What gets updated | Model weights and population of candidates |
| When it runs | During training |
| Validity | Approximate physics (simulation, not lab results) |
| Evidence | Experiments reported |
| Lineage | Original system |