MIT develops AI framework for stable material design

MIT researchers have announced a new AI-based framework, CrysVCD, that improves the stability of generated crystal materials while preserving desired properties.

• The approach uses a language model first to create chemically valid formulas, then a diffusion model to generate the crystal structure, achieving 68 percent mechanical stability and 85 percent metastability in fine-tuned tests.
• The paper was published in Nature Computational Science today, and the researchers say the method can support materials for high thermal conductivity, including applications in semiconductor industry and data center cooling; the work was supported by the U.S. Department of Energy, National Science Foundation, MathWorks Engineering Fellowship, and U.S. Defense Threat Reduction Agency.