SLayerGen: Crystal Generative Model for All Space and Layer Groups
SLayerGen is an innovative generative model tailored for crystals, capable of processing both space and layer groups, thus filling a void left by current models that focus solely on space groups. Layer groups are essential for describing diperiodic materials, such as 2D superconductors, thin film semiconductors, and catalytic surfaces, which exhibit periodicity in two dimensions while being aperiodic in the third. This model employs a coarse-to-fine methodology, incorporating discrete autoregressive lattice generation, transformer-based autoregressive sampling of Wyckoff positions, elements, and the count of symmetrically unique atoms, along with equivariant diffusion of atomic coordinates. The diffusion aspect rectifies an inconsistency found in previous studies concerning hexagonal groups. SLayerGen aims to enhance the exploration of bulk and diperiodic materials.
Key facts
- SLayerGen is a generative model for crystals invariant to any space or layer group.
- Layer groups describe diperiodic materials such as 2D superconductors, thin film semiconductors, and catalytic surfaces.
- Existing crystal generative models only consider space groups, not layer groups.
- SLayerGen uses coarse-to-fine discrete autoregressive lattice generation.
- It employs transformer-based autoregressive sampling of Wyckoff positions, elements, and numbers of symmetrically unique atoms.
- The model includes space or layer group equivariant diffusion of atomic coordinates.
- The diffusion component corrects an inconsistency in prior work arising from hexagonal groups.
- SLayerGen aims to accelerate discovery of bulk and diperiodic materials.
Entities
—