MIT's CrysVCD Framework Helps AI Design Materials That Are Actually Stable

Anyone with a large enough AI model can generate millions of new material designs in minutes. But that has not led to a big leap in the new materials used in products like computer chips and rockets. One reason: current models don't reliably factor in whether a material is chemically stable. Unstable materials are not useful in the real world, so industries spend huge computing budgets screening them out, often keeping only a tiny fraction of what was generated.

MIT researchers have developed a framework that works at the very start of the material-generation process instead. It is called CrysVCD, short for crystal generator with valence-constrained design. Before the expensive generation step even begins, the framework checks that every design satisfies key rules of chemistry about the electrons around a material's atoms. The work was published in Nature Computational Science.

CrysVCD combines an AI diffusion model, the kind commonly used to generate images, with a language model. First, the language model produces a chemically valid formula. Then the diffusion model uses that formula to generate the atomic structure of the crystal. “Diffusion for typical material generation is a slow process — you can think of it like 1,000 steps to create one material,” said Weiliang Luo, an MIT doctoral student. “In contrast, when our model is used in the beginning, you can think of it like five steps.”

Using CrysVCD, several commonly used material models met chemistry's valence-shell rules more often, and achieved high lattice-dynamics stability, a stringent stability test, in nearly 70% of computational material generations. When fine-tuned on stability metrics, the approach produced crystalline materials that achieved 68% mechanical stability and 85% metastability, a measure of whether a material stays in a stable state when undisturbed. The researchers say this beats screening materials only after they are generated.

The team then used CrysVCD to generate material candidates with high thermal conductivity and materials that polarize easily in an electric field, both useful for the semiconductor industry and for data centers. “If material-generating models are like DVDs, we are like the DVD player,” said Mingda Li, an MIT associate professor of nuclear science and engineering. “You can plug this into any kind of model.”