Crystal structure generative modeling based on diffusion probabilistic models and variational autoencoder

We have developed a machine learning model for generating crystal structures based on variational autoencoder (VAE) and diffusion probabilistic models. By adopting the crystal diffusion VAE [1], we show that physically plausible crystal structures can be generated through two neural networks: the decoder to generate lattice parameters and the number of atoms, and the diffusion network to generate the atomic coordinates by denoising the coordinates from random noises. As a result, the qualities of crystal structures generated by our model are statistically comparable to those generated by the crystal diffusion VAE. We benchmark the results with the density functional theory, revealing that the energy differences between relaxed and generated structures are, on average, 0.4 eV/atom. Our work shows a promising data-driven scheme of crystal structure generation that can well approximate true ground-state configurations.

Reference

[1] Tian Xie, Xiang Fu, Octavian-Eugen Ganea, Regina Barzilay, and Tommi Jaakkola, "Crystal Diffusion Variational Autoencoder for Periodic Material Generation", International Conference on Learning Representations (2022).