Inverse Design of Next-Generation Superconductors Using Data-Driven Deep Generative Models.

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Discovering new superconductors with high critical temperatures (Tc) is crucial but hindered by high computational and experimental costs.
• Traditional materials screening methods are often funnel-like, limiting the scope of discovery.
• Advanced computational tools are needed to accelerate the search for next-generation superconducting materials.

Purpose of the Study:

• To present a novel diffusion model for generating new superconductors with unique structures and chemical compositions.
• To leverage artificial intelligence and existing databases for efficient materials discovery.
• To enable the inverse design of materials with desired superconducting properties.

Main Methods:

• Utilized a crystal diffusion variational autoencoder (CDVAE) and an atomistic line graph neural network (ALIGNN) pretrained model.
• Trained the diffusion model on a dataset of approximately 1000 superconducting materials from density functional theory (DFT) calculations.
• Employed the Joint Automated Repository for Various Integrated Simulations (JARVIS) superconducting database.

Main Results:

• Generated 3000 new potential superconductor structures using the diffusion model.
• Screened generated structures with the pretrained ALIGNN model, identifying 61 promising candidates.
• Validated the top candidates through density functional theory (DFT) calculations, demonstrating a high success rate.

Conclusions:

• The developed diffusion model effectively generates novel superconductor candidates with high success rates.
• This AI-driven approach offers a powerful alternative to traditional materials screening, enabling inverse design.
• The methodology significantly reduces the costs associated with discovering high-Tc superconductors.