FlowMM
Riemannian flow matching on crystal flat tori and symmetric lattice manifolds with stability computed by CHGNet and DFT relaxations on generated samples.
The loop
- Proposes
A Riemannian flow-matching network integrates geodesic vector fields on the periodic torus of fractional coordinates and SPD lattice representations.
- Judged by
Symmetry-aware structural matchers, CHGNet force-field relaxations, and VASP DFT calculations evaluate thermodynamic stability (E_hull) on generated samples.
- Steers
DFT and CHGNet relaxations compute stability metrics on generated samples; the flow is trained on a fixed dataset.
Headline result
Achieved state-of-the-art thermodynamic stability and 3x faster integration steps than diffusion baselines on MP-20 and MPTS-52.
What this does not prove
Like other crystal generators, requires downstream DFT phonon calculations to confirm dynamic vibrational stability at finite temperature.
How it generalizes the harness
Riemannian flow matching on crystal manifolds generates candidate lattices whose stability is computed afterwards with CHGNet and DFT relaxations.
Classification
| What it proposes | Crystal structure |
|---|---|
| Generative model | Riemannian flow-matching crystal network |
| Algorithmic authority | Crystallographic manifold constraints and VASP DFT evaluator |
| Feedback | Score |
| What gets updated | Population of candidates |
| When it runs | At inference time |
| Validity | Approximate physics (simulation, not lab results) |
| Evidence | Code released |
| Lineage | Successor |