Technology

CDVAE

An SE(3)-invariant generative model that designs stable, periodic crystal structures by coupling variational autoencoders with diffusion processes.

Developed to bypass the limitations of traditional, grid-based material generation, the Crystal Diffusion Variational Autoencoder (CDVAE) targets the direct design of periodic material structures. The framework encodes a crystal's periodic lattice alongside its atomic coordinates and types into a continuous latent space. During decoding, a diffusion process iteratively updates atomic positions and species to satisfy physical bonding preferences, outputting stable structures while respecting translation, rotation, and permutation invariances. Tested on standard benchmarks like MP-20 and Carbon-24, CDVAE enables researchers to reconstruct structures, generate realistic materials, and optimize specific target properties via latent-space search.