Poster
in
Workshop: AI4Mat-2024: NeurIPS 2024 Workshop on AI for Accelerated Materials Design
WyckoffTransformer: Generation of Symmetric Crystals
Nikita Kazeev · Ruiming Zhu · Romanov Ignat · Andrey Ustyuzhanin · Shuya Yamazaki · Wei Nong · Kedar Hippalgaonkar
Keywords: [ Wyckoff position ] [ autoregressive model ] [ generative model ] [ machine learning ] [ material design ] [ Transformer ]
We propose WyckoffTransformer, a generative model for inorganic materials that takes advantage of the high order symmetry present in most known crystals. Wyckoff positions, a mathematical object from space group theory, is used as the basis for an elegant, compressed, and discrete structure representation. To model the distribution we develop a permutation–invariant autoregressive model based on Transformer. Our experiments demonstrate that Wyckoff Transformer has better performance compared to the baseline in generating novel stable structures conditioned on the space group symmetry, while also having competitive metric values when compared to a model not conditioned on space group symmetry.