Generative Models for Crystalline Materials.

Area of Science:

• Condensed matter physics and materials science.
• Computational materials science.
• Machine learning applications in scientific discovery.

Background:

• Understanding material structure-property relationships is crucial for scientific advancement.
• Machine learning (ML) is increasingly vital for accelerating materials discovery and design.
• Recent focus has shifted from screening to generative models for crystal structure prediction.

Purpose of the Study:

• To review the current landscape of generative models for crystal structure prediction and de novo generation.
• To analyze crystal representations, generative model types, and their limitations.
• To guide experimental scientists and ML specialists in inverse materials design.

Main Methods:

• Examination of various crystal structure representations.
• Outline and evaluation of different end-to-end generative models for crystal design.
• Analysis of strengths and limitations of current generative approaches.

Main Results:

• Generative models offer powerful capabilities for de novo crystal structure design.
• Key considerations include crystal representation, model architecture, and experimental validation.
• Emerging areas include modeling defects, disorder, and synthetic constraints.

Conclusions:

• Generative modeling is transforming inverse materials design and discovery.
• The review provides insights into model selection, evaluation, and future research directions.
• Bridging ML specialists and experimental scientists is key for practical application.