3-D Inorganic Crystal Structure Generation and Property Prediction via Representation Learning

Area of Science:

• Materials Science
• Artificial Intelligence
• Crystallography

Background:

• Generative models excel at creating novel content like images and text.
• Applying generative deep learning to discover stable, novel 3-D crystal structures across material classes remains a challenge.
• Existing methods have not yet yielded reliable 3-D crystal structure generation for materials science.

Purpose of the Study:

• To develop a generative deep-learning pipeline for designing geometrically optimized 3-D crystal structures.
• To predict eight target properties simultaneously for newly generated materials.
• To demonstrate the pipeline's generalizability across multiple material classes.

Main Methods:

• An autoencoder-based generative deep-representation learning pipeline was employed.
• The system was trained to generate and optimize 3-D crystal structures.
• Generated structures were validated against electronic-structure calculations.

Main Results:

• Novel, stable, and geometrically optimized 3-D crystal structures were successfully generated.
• The pipeline demonstrated high generality by creating materials from binary alloys, ternary perovskites, and Heusler compounds.
• Generated materials were confirmed as valid through comparison with established computational methods.

Conclusions:

• The developed autoencoder-based pipeline effectively generates novel and optimized 3-D crystal structures.
• This approach offers a powerful tool for accelerating the discovery of new functional materials.
• The system's success across diverse material classes highlights its broad applicability in materials design.