CrysLDM: Latent Diffusion Model for Crystal Material Generation

Generating new crystal materials with desirable chemical properties has long been a challenging task. Existing diffusion models operate in feature space, requiring complex diffusion architectures to model the joint distribution of atom types, coordinates, and lattice structures. This complexity increases the number of diffusion steps, leading to higher training and sampling costs. In this work, we aim to generate novel crystal materials within a time- and resource-constrained setup, where existing models are not well-suited. To address this, we propose CrysLDM, a novel latent diffusion model for 3D crystal materials, which integrates a variational autoencoder (VAE) and a diffusion model. The VAE encoder maps 3D crystal structures into a latent space, where the diffusion model operates. Since CrysLDM leverages a smooth, lower-dimensional latent space, it simplifies the generative process and accelerates both training and inference. Through extensive experiments on benchmark datasets and tasks, we show that CrysLDM generates stable and valid materials with quality comparable to state-of-the-art methods, while being significantly more efficient in terms of computational resources and time.