CrysVCD applies chemical rules before generating new materials

An MIT method filters compositions at an early stage to increase the proportion of stable structures proposed by generative models.

On August 26, MIT presented CrysVCD, a method that frames the generation of crystals by valence rules. A language model first proposes chemically admissible formulas; a scattering model then produces the corresponding atomic structures.

The researchers report nearly 70% dynamic stability in a computational evaluation, reducing the need to filter a large number of unstable candidates after generation. The method also makes it possible to target certain properties, such as thermal conductivity. These calculated criteria are used to guide the search for materials and do not replace their experimental manufacture and characterization.