Crystal structure prediction with host-guided inpainting generation and foundation potentials

Area of Science:

• Materials Science
• Computational Chemistry
• Crystallography

Background:

• Diffusion models struggle with generating symmetric crystal structures as unit cell size increases.
• Accurate crystal structure generation is crucial for materials discovery and design.

Purpose of the Study:

• To develop a novel framework, crystal host-guided generation (CHGGen), to improve the generation of symmetric crystal structures.
• To address the limitations of unconditional crystal structure generation methods.

Main Methods:

• Implemented a conditional generation approach using an inpainting method within the CHGGen framework.
• Optimized a fraction of atomic positions in a predefined, symmetrized host structure.
• Integrated inpainting structure generation with a foundation potential for structure optimization.

Main Results:

• The CHGGen framework significantly improves the success rate of generating symmetric crystal structures.
• Demonstrated higher fractions of symmetric structures generated using the inpainting method compared to unconditional generation in ZnS-P2S5 and Li-Si systems.
• Validated the practical significance for modifying crystal structures with partial occupancy or intercalation chemistry.

Conclusions:

• CHGGen offers a versatile and effective approach for generating symmetric crystal structures.
• The inpainting method facilitates seamless integration with other generative models, accelerating materials discovery.
• CHGGen advances the field of computational materials science by enabling more accurate and efficient structure generation.