Practical approaches for crystal structure predictions with inpainting generation and universal interatomic potentials (2504.16893v1)

Abstract: We present Crystal Host-Guided Generation (CHGGen), a diffusion-based framework for crystal structure prediction. Unconditional generation with diffusion models demonstrates limited efficacy in identifying symmetric crystals as the unit cell size increases. CHGGen addresses this limitation through conditional generation with the inpainting method, which optimizes a fraction of atomic positions within a predefined and symmetrized host structure. We demonstrate the method on the ZnS-P$_2$S$_5$ and Li-Si chemical systems, where the inpainting method generates a higher fraction of symmetric structures than unconditional generation. The practical significance of CHGGen extends to enabling the structural modification of crystal structures, particularly for systems with partial occupancy, surface absorption and defects. The inpainting method also allows for seamless integration with other generative models, providing a versatile framework for accelerating materials discovery.