Crystal structure prediction based on diffusion model and graph network optimization
Accurately and quickly predicting the crystal structures of unknown materials can significantly accelerate the discovery of new materials. However, stable crystal structures only exist at the global minimum of formation energy, making optimization based on density functional theory calculations or g...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IOP Publishing
2025-01-01
|
| Series: | JPhys Materials |
| Subjects: | |
| Online Access: | https://doi.org/10.1088/2515-7639/adeaed |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Accurately and quickly predicting the crystal structures of unknown materials can significantly accelerate the discovery of new materials. However, stable crystal structures only exist at the global minimum of formation energy, making optimization based on density functional theory calculations or graph neural networks (GNNs) surrogate models time-consuming due to the extensive search space. In this work, we propose a crystal structure prediction method called DiffOA, which combines a diffusion model with an optimization algorithm based on GNNs. By generating atomic coordinates through a diffusion process and minimizing the formation energy through an optimization process, DiffOA effectively reduces the searching space required for crystal structure prediction while respecting both the minimum energy principle and the symmetry of crystal structures. We evaluated our method on the prediction of crystal structures of 29 compounds and found that DiffOA achieves a speed three times faster while maintaining comparable performance to GNNs-based optimization method. The high efficiency and accuracy of DiffOA may open a new path for data-driven materials discovery. |
|---|---|
| ISSN: | 2515-7639 |