Scalable crystal structure relaxation using an iteration-free deep generative model with uncertainty quantification.

Area of Science:

• Computational molecular and materials science
• Artificial intelligence for science
• Materials discovery

Background:

• Determining equilibrium structures is essential for accurate property calculations in computational materials science.
• Traditional methods struggle with the scale of newly discovered complex crystal structures.
• Iterative processes in existing computational methods limit scalability for large-scale screening.

Purpose of the Study:

• Introduce DeepRelax, a novel deep generative model for rapid crystal structure relaxation.
• Address the scalability challenges of traditional computational methods for materials science.
• Enable fast and accurate prediction of equilibrium structures for large datasets.

Main Methods:

• Developed DeepRelax, a deep generative model learning equilibrium structural distributions.
• Implemented direct prediction of relaxed structures from unrelaxed ones, bypassing iterative processes.
• Integrated uncertainty quantification to enhance model trustworthiness.

Main Results:

• DeepRelax achieves millisecond-level relaxation times per structure.
• Demonstrated high accuracy and efficiency across five diverse material databases (oxides, Materials Project, 2D materials, van der Waals crystals, crystals with defects).
• Validated results against density functional theory calculations.

Conclusions:

• DeepRelax significantly accelerates computational workflows in materials science.
• Offers a robust, trustworthy, and scalable machine-learning solution for material discovery.
• Advances the application of artificial intelligence in scientific research and development.