Generative AI can dream up crystal structures by the thousand, but many of its proposals are chemically unstable, forcing laboratories to burn computing resources filtering them out and sometimes keeping only a small fraction. A team led by Associate Professor Mingda Li, with MIT colleagues Ju Li and Heather Kulik and collaborators at Oak Ridge National Laboratory and Michigan State University, has addressed the problem at source with CrysVCD, a framework published in Nature Computational Science.
CrysVCD works in two stages. A language model first proposes chemical formulas that obey electron-based valence rules, guaranteeing a plausible composition, and a diffusion model then generates the three-dimensional atomic structure for that formula. In testing, the approach produced stable materials roughly ten times more efficiently than generating first and screening afterwards; when tuned explicitly for stability, it achieved 68 percent mechanical stability and 85 percent metastability among its outputs.
The team also steered the generator towards targeted properties, including candidates with high thermal conductivity relevant to semiconductor design and data-centre cooling. Because the valence-checking stage can bolt onto existing and future generative models, the method could cut the computational cost of materials discovery for smaller laboratories and companies, though it performs best on highly ordered crystalline solids.