CDVAE
Crystal Diffusion Variational Autoencoder
Periodic SE(3)-invariant variational autoencoder and Langevin dynamics generator whose latent optimization is guided by a learned property predictor.
The loop
- Proposes
A periodic GNN decoder predicts unit-cell lattice parameters, stoichometry, and denoises atomic coordinates via annealed Langevin dynamics.
- Judged by
A learned property predictor guides latent optimization; validity checks (symmetry, bond distance, charge neutrality) and a separate learned energy model evaluate finished samples.
- Steers
Gradient ascent in latent space steers sampling toward lower formation energy.
Headline result
Generated valid periodic crystal structures across carbon, perovskite, and Materials Project benchmarks with high structural and compositional validity.
What this does not prove
Latent-space property optimization relies on a learned surrogate predictor.
How it generalizes the harness
Uses a periodic SE(3)-equivariant diffusion VAE rather than an LLM, steered by a learned property predictor and evaluated with validity checks and a learned energy model.
Classification
| What it proposes | Crystal structure |
|---|---|
| Generative model | Periodic SE(3)-equivariant diffusion VAE |
| Algorithmic authority | Crystallographic validity checker and a learned energy model |
| 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 | Original system |