CrystalDiT: A Diffusion Transformer for Crystal Generation
By: Xiaohan Yi , Guikun Xu , Xi Xiao and more
Potential Business Impact:
Designs new, stable materials faster and easier.
We present CrystalDiT, a diffusion transformer for crystal structure generation that achieves state-of-the-art performance by challenging the trend of architectural complexity. Instead of intricate, multi-stream designs, CrystalDiT employs a unified transformer that imposes a powerful inductive bias: treating lattice and atomic properties as a single, interdependent system. Combined with a periodic table-based atomic representation and a balanced training strategy, our approach achieves 9.62% SUN (Stable, Unique, Novel) rate on MP-20, substantially outperforming recent methods including FlowMM (4.38%) and MatterGen (3.42%). Notably, CrystalDiT generates 63.28% unique and novel structures while maintaining comparable stability rates, demonstrating that architectural simplicity can be more effective than complexity for materials discovery. Our results suggest that in data-limited scientific domains, carefully designed simple architectures outperform sophisticated alternatives that are prone to overfitting.
Similar Papers
MiAD: Mirage Atom Diffusion for De Novo Crystal Generation
Machine Learning (CS)
Finds new crystal materials by adding/removing atoms.
All-atom Diffusion Transformers: Unified generative modelling of molecules and materials
Machine Learning (CS)
Creates new molecules and materials with one tool.
LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generation
Machine Learning (CS)
Creates new, stable materials for science.