MIT introduced CrysVCD, moving chemical valence constraints to the front of materials generation

MIT researchers combined a language model with diffusion generation so chemically valid formulas come before crystal structures; nearly 70 percent of generated candidates passed a stringent lattice-dynamics stability test, with target properties including chip-cooling performance.