Crystal Structure Prediction Using Generative Adversarial Network with Data-Driven Latent Space Fusion Strategy.

Area of Science:

• Materials Science
• Computational Materials Design
• Crystallography

Background:

• Crystal structure prediction (CSP) is crucial for materials design but faces challenges with high-dimensional data.
• Generative Adversarial Networks (GANs) are powerful tools but suffer from issues like mode collapse.
• Existing methods require improvement for accurate and efficient prediction of complex crystal structures.

Purpose of the Study:

• To introduce a novel GAN-based model (GAN-DDLSF) for enhanced crystal structure prediction.
• To address the limitations of current GANs in materials science by optimizing latent space representation.
• To improve the accuracy and efficiency of predicting binary crystal structures, using gallium nitride (GaN) as a case study.

Main Methods:

• Developed a novel generative adversarial network model named GAN-DDLSF.
• Introduced a data-driven latent space fusion (DDLSF) sampling method to optimize GANs' latent space.
• Combined statistical properties of real crystal data with Gaussian distribution to mitigate mode collapse.
• Refined the generation mechanism for binary crystal structures, focusing on crystallographic features of GaN.

Main Results:

• Generated 9321 binary crystal structures for gallium nitride (GaN).
• Achieved 16.59% stable and 24.21% metastable structures, indicating high prediction accuracy.
• Demonstrated improved precision and efficiency in predicting GaN structures.
• Validated the GAN-DDLSF approach for materials discovery.

Conclusions:

• The GAN-DDLSF model with DDLSF sampling effectively enhances crystal structure prediction accuracy.
• The approach shows significant potential for the design and discovery of binary, ternary, and multinary materials.
• This work offers new methodologies for materials science research and applications in computational materials design.