Generating millions of new material designs with artificial intelligence takes minutes. Turning those designs into materials that actually work in computer chips, rockets, or data centers is harder, because most AI models do not reliably produce chemically stable structures, forcing companies to spend enormous computing budgets filtering out unusable candidates.
MIT researchers have built a framework that filters for stability at the start of the generation process rather than after it, cutting the cost of finding usable materials while still targeting specific properties. They call the approach CrysVCD, short for crystal generator with valence-constrained design. In a paper published in Nature Computational Science, the team reported that several widely used material-generation models, when paired with CrysVCD, met valence-shell rules more often and produced high lattice-dynamics stability — a stringent test — in nearly 70 percent of computational generations. The framework also supported materials with targeted properties such as high thermal conductivity and a high dielectric constant, both relevant to semiconductors.
How CrysVCD Works
The system works in two stages. A language model first produces chemically valid formulas by enforcing rules about the electrons surrounding each atom. A diffusion model then generates the corresponding atomic crystal structure. The diffusion step normally takes around 1,000 steps to build one material; the new approach front-loads the chemistry checks, so the generation runs in roughly 5 steps.
“If material-generating models are like DVDs, we are like the DVD player,” says Mingda Li, an associate professor of nuclear science and engineering. “You can plug this into any kind of model, not only existing diffusion models but also future models, where people can’t generate enough stable materials, and it can improve stability.”
The Cost of Validation
The cost of validating stability has become a bottleneck. Mouyang Cheng, a materials science and engineering doctoral student, estimates that validation accounts for about 90 percent of the computing cost of producing usable materials, a process that can take weeks or months. Large companies with deep compute budgets can absorb that expense. Smaller labs and academic groups often cannot, which limits who can contribute to the field.
“In academia, where we have fewer resources, I think we can still achieve strong performance with smarter designs and other approaches,” says Heather Kulik, the Lammot du Pont Professor of Chemical Engineering. “Generating a model and then down-selecting for stability is inefficient. There’s a high computational cost. But if we put a language model in the beginning of the process to constrain the generation, you can significantly enhance the ratio of stable materials generated.”
Stability and Performance Results
In tests, the MIT approach produced more stable materials about an order of magnitude more efficiently than pipelines that screen after generation. When fine-tuned for stability, the system generated crystalline materials with 68 percent mechanical stability and 85 percent metastability — a measure of whether a material remains stable when undisturbed.
The team then asked the system to design materials with high thermal conductivity and easy polarization in an electric field. Both properties matter for semiconductors and for cooling data centers, where roughly 30 percent of energy use goes to heat removal.
“These are materials useful for the semiconductor industry and high thermal conductivity materials relevant to data center cooling,” says Ju Li, the Carl Richard Soderberg Professor in Power Engineering. “In principle, you could also use this to create other properties, but thermal conductivity has become really important for cooling data centers. There’s been a huge increase in energy use in that industry, and 30 percent of that energy goes to cooling. The industry needs materials with high thermal conductivity to more efficiently remove the heat.”
Scope and Limitations
CrysVCD is built for solid, highly ordered structures and does not cover every class of material. Within its scope, however, the researchers say it can target stability and performance at the same time, a combination that has been difficult to achieve.
“We are not just generating stable materials, we’re also prioritizing performance,” Cheng says. “Any time you have two goals, achieving those goals with anything over 50 percent is hard in this field. In the past, people might have a goal for specific properties and not stability, or vice-versa, and get a single-digit percentage of materials that fit their goal.”
Joining Mingda Li on the paper are Mouyang Cheng, Weiliang Luo, Hao Tang, Bowen Yu, Yongqiang Cheng of Oak Ridge National Laboratory, Weiwei Xie of Michigan State University, Ju Li, and Heather Kulik. The work was supported by the U.S. Department of Energy, a Mathworks Engineering Fellowship, the National Science Foundation, and the U.S. Defense Threat Reduction Agency.