MIT researchers have developed a framework that can be applied at the start of the materials generation process to dramatically improve the stability rate of AI-generated materials while still achieving targeted properties. The approach, detailed today in Nature Computational Science, works by ensuring every design satisfies key rules of chemistry relating to the electrons around atoms before the expensive generation step begins, addressing a major bottleneck that has limited AI's practical impact on materials discovery.
Anyone with a large enough AI model can now generate millions of new material designs in minutes. But that hasn't translated into a surge of new materials being used in products like computer chips and rockets. One reason for the gap: current models don't reliably factor in the chemical stability of the materials they generate, and unstable materials aren't useful in the real world. Industries have been forced to allocate huge computational budgets to screening out unstable materials, in some cases leaving behind only a tiny fraction of usable options.
The MIT team calls their approach "crystal generator with valence-constrained design," or CrysVCD. In testing, it allowed several commonly used material models to meet valence shell rules more often, achieving high lattice-dynamics stability - a stringent stability test - in nearly 70 percent of computational material generations. The researchers also showed the approach could support the creation of materials with specific desired properties, like high thermal conductivity or high dielectric constant, which matters for computer chips and data centers.
Fixing the stability bottleneck
Computational approaches to materials design have existed for decades, but recent advances in AI have intensified interest in their potential. Of particular interest are models that start with a desired material property and work backward to deliver a material that achieves that goal. Some use diffusion techniques, commonly used to generate images, while others use large language models similar to those powering ChatGPT and Claude. Both approaches struggle to ensure their generations achieve chemical stability or follow fundamental principles about how chemicals interact.
The conventional fix has been to add a filtering layer on top of the generative process to remove unstable materials. That's expensive. "It's becoming easy to generate the material structure," says Mouyang Cheng, an MIT doctoral student in materials science and engineering. "But the validation process, especially the part where you test the stability, has a huge computational cost. It's something like 90 percent of the computational cost for creating usable materials, and it can take weeks or months."
Large companies with big computing budgets can afford those processes, but many small companies and research labs can't. "In academia, where we have fewer resources, I think we can still achieve strong performance with smarter designs and other approaches," says Heather Kulik, MIT's 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."
CrysVCD combines AI diffusion models with a language model. In the first stage, the language model produces chemically valid formulas. In the second stage, the diffusion model uses that formula to generate the corresponding atomic structure of the crystal material. The approach is model-agnostic, which is central to its design.
"If material-generating models are like DVDs, we are like the DVD player," says Mingda Li, associate professor of nuclear science and engineering at MIT. "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."
"Diffusion for typical material generation is a slow process - you can think of it like 1,000 steps to create one material," says Weiliang Luo, an MIT doctoral student in chemistry. "In contrast, when our model is used in the beginning, you can think of it like five steps. It allows you to screen out the unstable materials to generate higher quality materials. And it works with any models generating materials," adds Hao Tang, a recent MIT graduate in materials science and engineering.
The researchers demonstrated that their approach created stable materials an order of magnitude more efficiently than methods relying on post-generation screening. When fine-tuned on stability metrics, it produced crystalline materials that achieved 68 percent mechanical stability and 85 percent metastability, which measures whether a material stays in a stable state when undisturbed.
Targeting properties that industry needs
The team then used their approach to generate material candidates with high thermal conductivity and easy polarization in an electric field - properties directly relevant to the semiconductor industry and data center cooling. "These are materials useful for the semiconductor industry and high thermal conductivity materials relevant to data center cooling," says Ju Li, MIT's 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."
The new approach doesn't work with every kind of material - it works best with solid structures that have highly ordered internal arrangements. Still, it could be used to generate stable new crystalline materials with a range of important properties.
"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."
The work connects to broader developments in Generative AI and LLM applications, where the focus is shifting from raw generation capability to reliability and domain-specific constraints. The MIT framework is one example of how researchers are making generative models more useful for scientific discovery.
Why this matters for science and research professionals
For researchers working in materials science, chemistry, and related fields, the practical takeaway is computational cost reduction. The approach eliminates the need for expensive downstream screening, which currently consumes the vast majority of computing resources in materials generation workflows. "This will save huge computation costs and time by removing downstream selection requirements," Mingda Li says. "That will help not only large efforts that generate hundreds of millions of materials, but also smaller research groups with targeted applications."
For those applying AI for Science & Research, the framework represents a shift in how generative models are deployed: instead of generating broadly and filtering afterward, the constraints are built into the generation process itself. That distinction matters for anyone designing workflows that depend on AI-generated candidates, whether for energy storage, electronics, or thermal management.
The work was supported in part by the U.S. Department of Energy, a Mathworks Engineering Fellowship, the National Science Foundation, and the U.S. Defense Threat Reduction Agency.
Your membership also unlocks:
