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

This study develops the Crystal Host-Guided Generation (CHGGen) framework, which applies inpainting methods from computer vision to crystal structure prediction. Instead of generating complete structures unconditionally, CHGGen optimizes atomic positions within predefined, symmetrized host lattices, thereby improving long-range crystallographic order.

The model is integrated with a foundation potential for structure relaxation, enabling both structure discovery and structural modification of systems with partial occupancy or intercalation chemistry. Benchmarking on ZnS–P₂S₅ and Li–Si chemical systems shows that CHGGen produces a higher fraction of symmetric structures than unconditional diffusion models.

This conditional generative approach provides a versatile, modular tool for accelerating materials discovery, particularly in domains like battery materials.

This study applies an SE(3)-equivariant diffusion model, pre-trained on the Materials Project database, to efficiently predict intercalant positions in host structures for energy storage materials design using the inpainting method.

• Introduced CHGGen, a host-guided inpainting framework for crystal structure generation.
• Overcame the locality bias of graph neural network diffusion models by enforcing host-symmetry constraints.
• Achieved higher success in generating symmetric and chemically plausible structures for complex polyanion systems.
• Demonstrated versatility for structural modification in partial occupancy and intercalation contexts.
• Integrated with foundation potentials (e.g., CHGNet) to validate structural stability with low interatomic forces and minimal energy deviation.

Download paper here