Siamese foundation models for crystal structure prediction.

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence in Materials Discovery

Background:

• Predicting crystal structures from chemical composition is crucial for materials discovery but challenging due to complex 3D geometries.
• Existing computational methods can be slow and struggle with complex or novel structures.

Purpose of the Study:

• To introduce Diffusion-based crystAl Omni (DAO), a novel pretrain-finetune framework for enhanced crystal structure prediction.
• To leverage Siamese foundation models for structure generation and energy prediction to improve accuracy and speed.

Main Methods:

• Developed DAO, integrating a structure generator and an energy predictor within a Siamese network framework.
• Pretrained the generator using a two-stage pipeline on a large dataset, utilizing the predictor for configuration relaxation and guided sampling.
• Validated the framework on established benchmarks and real-world superconductor materials.

Main Results:

• Pretraining significantly improved performance across various backbone architectures on standard benchmarks.
• Synergistic interaction between the generator and predictor mutually enhanced both model components.
• Achieved a 100% match rate and low atomic-position error for Cr6Os2, outperforming DFT-based methods by over 2000x per iteration.

Conclusions:

• DAO demonstrates superior performance and efficiency in crystal structure prediction compared to conventional computational approaches.
• The framework shows significant potential for accelerating the discovery of novel materials, including complex superconductors.
• The integrated generator-predictor approach offers a powerful new paradigm for computational materials science.