Self-supervised generative models for crystal structures.

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Advancements in natural language processing inspire new machine learning approaches.
• Generating reliable inorganic crystal structures is crucial for materials discovery.
• Predicting material properties requires efficient and accurate models.

Purpose of the Study:

• To develop a unified platform for generative models of inorganic crystal structures.
• To enable efficient adaptation to downstream tasks like material property prediction.
• To enhance the reliability evaluation of generated structures during training.

Main Methods:

• Utilized self-supervised learning and equivariant graph neural networks.
• Employed a generative adversarial network (GAN) with a cost-effective discriminator for reliability evaluation.
• Demonstrated model utility in optimizing crystal structures and grouping chemically similar elements.

Main Results:

• Successfully generated inorganic crystal structures and predicted material properties.
• Significantly enhanced model performance through GAN-based reliability evaluation.
• Showcased the model's ability to optimize structures and understand crystal formation without external data.

Conclusions:

• The developed platform offers a novel perspective on machine learning in material science.
• Generative models can effectively aid in understanding inorganic crystal formation and properties.
• This work paves the way for further exploration of AI in discovering and designing new materials.