CrystalDiT: A Diffusion Transformer for Crystal Generation

TL;DR

CrystalDiT introduces a simplified diffusion transformer model for crystal structure generation, achieving state-of-the-art results by leveraging a unified architecture and atomic representations, outperforming complex models in stability, uniqueness, and novelty.

Contribution

The paper proposes a unified transformer architecture for crystal generation that challenges the trend of complex designs, demonstrating superior performance with a simpler model.

Findings

Achieves 8.78% SUN rate on MP-20, outperforming recent methods.

Generates 63.28% unique and novel crystal structures.

Simpler architecture outperforms complex models in data-limited scenarios.

Abstract

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 8.78% SUN (Stable, Unique, Novel) rate on MP-20, substantially outperforming recent methods including FlowMM (4.21%) and MatterGen (3.66%). 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.