Unconditional Crystal Generation

This task evaluates the ability of generative models to sample valid and realistic crystal structures.

The training dataset, MP20 (Xie et al., 2022), contains 45,231 metastable crystal structures from the Materials Project (Jain et al., 2013), each with up to 20 atoms and spanning 89 different element types.

Following prior work (Xie et al., 2022), we sample 10,000 crystals and compute validity, stability, uniqueness and novelty rates using density functional theory (DFT). Data is generally split following prior work (Xie et al., 2022) to ensure fair comparisons.