MIT researchers have developed CrysVCD, an AI framework designed to apply chemical constraints before generating complete crystal structures. The system uses a language model to produce chemically valid material formulas and a diffusion model to create atomic structures. In testing, nearly 70% of generated materials achieved lattice-dynamics stability. The researchers also reported mechanical stability of 68% and metastability of 85% after fine-tuning, with potential applications in semiconductor materials, thermal management and data-centre cooling.
