Data-Driven Score-Based Models for Generating Stable Structures with Adaptive Crystal Cells.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Discovering new functional and stable materials is complex and challenging.
• Generating crystal structures presents unique difficulties due to periodicity and symmetry constraints.

Purpose of the Study:

• To develop a machine learning generative model for creating novel crystal structures with desired properties like chemical stability and composition.
• To address the challenges of crystal structure generation, including lattice and atomic position determination.

Main Methods:

• Adapted score-based probabilistic models using annealed Langevin dynamics for crystal generation.
• Introduced a novel approach where the crystal lattice is learned and generated alongside atomic positions.
• Utilized a multigraph crystal representation that respects symmetry constraints.

Main Results:

• The model successfully generates new candidate crystal structures across various chemical systems and crystal groups without retraining.
• The multigraph representation offers computational advantages and improves the quality of generated structures.
• Demonstrated the model's capability through a comparative analysis with existing generative models.

Conclusions:

• The proposed machine learning model offers a powerful and flexible method for de novo crystal structure generation.
• This approach advances the design of novel materials with tailored properties by overcoming traditional limitations.