Design of New Inorganic Crystals with the Desired Composition Using Deep Learning.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Discovering new solid-state materials is crucial for technological advancement.
• Generative models, particularly deep learning, show promise for crystal structure prediction.
• Existing models often struggle with specific elemental compositions or generate random structures.

Purpose of the Study:

• To develop a machine learning model capable of generating crystal structures with desired compositions.
• To explore the potential of crystal diffusion variational autoencoders for novel material discovery.
• To generate and validate new stable inorganic materials.

Main Methods:

• Developed a crystal diffusion variational autoencoder for generating structures with target compositions.
• Generated crystal structures for 14 compositions across three material types.
• Utilized density functional theory (DFT) calculations for structure stabilization and energy validation.
• Compared formation energies with traditional atom substitution and other generative models.

Main Results:

• Successfully generated crystal structures for 14 compositions, with 8 being novel.
• Identified the most stable structures in the existing database for 13 out of 14 compositions.
• Generated 205 unique new crystal materials with energy above hull <100 meV/atom.
• Achieved lower formation energies compared to traditional methods for most compositions.

Conclusions:

• The developed model effectively generates stable inorganic materials with specified compositions.
• This approach demonstrates significant potential for designing a wide range of new materials.
• The method shows promise for stable application across various scientific and industrial fields.