SLayerGen: A Generative Model for Diperiodic and Bulk Crystals
Researchers have introduced SLayerGen, a novel generative model designed to accelerate the discovery of both bulk periodic materials and diperiodic systems, such as 2D superconductors, thin film semiconductors, and catalytic surfaces. Unlike existing models that primarily focus on bulk crystals invariant under space groups, SLayerGen accounts for layer group symmetries, which are critical for materials aperiodic along one lattice direction. The model employs a coarse-to-fine discrete autoregressive approach for lattice generation, transformer-based sampling for atomic positions and elements, and equivariant diffusion for atomic coordinates. Notably, the authors corrected a loss inconsistency in prior work related to non-orthogonal hexagonal groups. To support this advancement, the team assembled filtered datasets for monolayers and bilayers, proposed new evaluation metrics, and developed novel representations for layer group symmetries. Experimental results indicate that SLayerGen achieves consistent performance gains over bulk-only models for de novo generation of diperiodic materials and remains competitive when trained jointly on diverse material types, marking a significant step forward in computational materials science and AI-driven discovery.
Wire timeline
SLayerGen: A Generative Model for Diperiodic and Bulk Crystals
Researchers have introduced SLayerGen, a novel generative model designed to accelerate the discovery of both bulk periodic materials and diperiodic systems, such as 2D superconductors, thin film semiconductors, and catalytic surfaces. Unlike existing models that primarily focus on bulk crystals invariant under space groups, SLayerGen accounts for layer group symmetries, which are critical for materials aperiodic along one lattice direction. The model employs a coarse-to-fine discrete autoregressive approach for lattice generation, transformer-based sampling for atomic positions and elements, and equivariant diffusion for atomic coordinates. Notably, the authors corrected a loss inconsistency in prior work related to non-orthogonal hexagonal groups. To support this advancement, the team assembled filtered datasets for monolayers and bilayers, proposed new evaluation metrics, and developed novel representations for layer group symmetries. Experimental results indicate that SLayerGen achieves consistent performance gains over bulk-only models for de novo generation of diperiodic materials and remains competitive when trained jointly on diverse material types, marking a significant step forward in computational materials science and AI-driven discovery.
cs.AI updates on arXiv.org