HOME > CDVAE

CDVAE

CDVAE is a diffusion-based variational autoencoder for crystalline materials. It learns from existing crystal structures and can generate novel periodic structures, as well as optimize for target properties in latent space. The project provides benchmark datasets (Perov-5, Carbon-24, MP-20) and training/evaluation scripts for crystal generation and property conditioning.

Information
Case / Article
Learning Content
Review
Developer’s Voice

Information

Official site	https://github.com/txie-93/cdvae
Openness	★★★
License	MIT License
Core Developers	Developers: Tian Xie, Xiang Fu, and collaborators
Related Papers	Tian Xie, Xiang Fu, et al., “Crystal Diffusion Variational AutoEncoder (CDVAE),” arXiv:2110.06197 (2021). https://arxiv.org/abs/2110.06197
Other	Description: CDVAE combines diffusion modeling with variational autoencoding to generate periodic materials while respecting crystal symmetry and lattice constraints. Its dataset suite targets three standard generation settings: composition-fixed carbon structures, perovskites with shared structure prototypes, and general inorganic materials capped at 20 atoms per cell. The codebase uses a Hydra configuration system, PyTorch Lightning training loops, and evaluation scripts tailored to structure reconstruction, generative validity, and property-optimization tasks. For data-driven materials science, CDVAE is a strong reference implementation for generative crystal modeling and dataset curation best practices.