Density of States-Intermediated Crystal Generation for Material Inverse Design

Sujong Lee, Jaewon Bae, Seungyun Shin, Heesung Choi, Chanyoung Park

Track: Track 1: Original Research/Position/Education/Attention Track

Keywords: AI4Materials, Crystal Generation, Inverse Design

Abstract:

Inverse design of crystalline materials often relies on property-conditional generation that directly maps target properties to crystal structures.
However, existing approaches ignore physical priors and fail to leverage the intrinsic relationships between physical quantities and crystal structures.
To address this, we present \emph{DOS-intermediated crystal generation}, a novel inverse design framework for crystalline materials with a physics-grounded intermediate variable.
We construct a two-stage pipeline: (1) property-conditional DOS generation, where a Masked Diffusion Model produces plausible DOS via prior token assignment and sample rejection; and (2) DOS-conditional crystal generation, where a fine-tuned MatterGen backbone produces crystals matching the generated DOS.
Empirically, our framework achieves robust performance on multi-property conditional generation and addresses diverse practical materials discovery scenarios beyond the reach of direct property conditioning.
We establish the first baseline for DOS-intermediated multi-property inverse design, validating this paradigm and providing a foundation that complements direct-generation models.

Submission Number: 162