Scaling deep learning for materials discovery.

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Traditional inorganic crystal discovery relies on costly trial-and-error methods, hindering rapid technological advancement.
• Deep learning models have demonstrated significant predictive power in various scientific domains, suggesting potential for materials science applications.

Purpose of the Study:

• To develop and apply large-scale graph networks for significantly enhancing the efficiency and scope of inorganic crystal discovery.
• To identify novel stable crystal structures beyond human chemical intuition and expand the known materials landscape.

Main Methods:

• Training graph networks on a dataset of 48,000 known stable crystals.
• Utilizing scaled deep learning to predict and discover new stable crystal structures.
• Performing hundreds of millions of first-principles calculations to validate stability and properties.

Main Results:

• Achieved an order-of-magnitude improvement in materials discovery efficiency.
• Identified 2.2 million new stable crystal structures, many previously unknown.
• 736 of the discovered stable structures have been experimentally validated.
• Developed highly accurate learned interatomic potentials for molecular dynamics simulations.

Conclusions:

• Large-scale graph networks represent a paradigm shift in materials discovery, overcoming limitations of traditional methods.
• The discovered materials offer vast potential for applications in clean energy, information processing, and beyond.
• The developed computational framework and discovered materials accelerate scientific breakthroughs and technological innovation.