Artificial intelligence can now generate millions of potential new materials in minutes. But there is a major problem: most of those designs may never survive the journey from a computer model to the real world.An AI system can suggest materials for faster computer chips, better rocket components or more efficient data centres, but if those materials are chemically unstable, researchers may have to spend weeks or even months — and enormous computing resources — simply screening them out. Now, researchers at MIT have developed a tool designed to tackle that problem before the expensive material-generation process even begins.
The problem with asking AI to invent new materials
Computational material design is not new, but artificial intelligence has dramatically expanded the scale at which scientists can search for new possibilities. Modern models can begin with a desired property — such as high thermal conductivity — and work backwards to suggest a material that might possess it.The difficulty comes after generation.As MIT News explained while reporting the research, existing AI systems do not always account for whether the materials they produce obey fundamental chemical rules or remain stable. As a result, researchers can generate vast numbers of candidates, only to discover that a tiny fraction are actually usable.“It’s becoming easy to generate the material structure,” Mouyang Cheng, one of the researchers, told MIT News. “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.”
MIT’s answer: Check the chemistry before generating the material
The new framework is called CrysVCD, short for crystal generator with valence-constrained design.Instead of generating millions of possible materials and then checking which ones are chemically stable, the MIT researchers built a system that applies key chemical constraints at the beginning.The approach focuses on rules governing the electrons surrounding atoms, known as valence constraints. By ensuring that a proposed material formula satisfies these fundamental chemical requirements before the full atomic structure is generated, the system can reduce the number of unstable designs entering the computational pipeline.Associate professor Mingda Li offered a simple analogy to MIT News: “If material-generating models are like DVDs, we are like the DVD player.”The idea, he said, is that CrysVCD can potentially be used alongside different material-generating models rather than functioning as a replacement for them.
From 1,000 steps to just five?
The researchers combined two forms of AI. A language model first generates chemically valid material formulas, after which a diffusion model produces the atomic structure of the crystalline material.According to MIT News, this could make the process dramatically more efficient.“Diffusion for typical material generation is a slow process — you can think of it like 1,000 steps to create one material,” researcher Weiliang Luo said.By contrast, Hao Tang explained that introducing the new system at the beginning of the process could reduce the initial screening process to something closer to “five steps”, while improving the quality of the generated candidates.
Nearly 70% passed a stringent stability test
In the study, published in Nature Computational Science, the researchers tested CrysVCD with commonly used material-generation models.The approach achieved high lattice-dynamics stability — a stringent computational test of whether a material structure is stable — in nearly 70% of generated materials.When fine-tuned using stability metrics, the system produced crystalline materials with 68% mechanical stability and 85% metastability.The researchers also used the framework to search for materials with specific useful properties, including high thermal conductivity and a high dielectric response.
Why this could matter for chips and AI data centres
These properties have applications in the semiconductor industry and data centres, where heat management has become an increasingly important engineering challenge.“Thermal conductivity has become really important for cooling data centers,” Ju Li told MIT News. “There’s been a huge increase in energy use in that industry, and 30 percent of that energy goes to cooling.”The researchers believe the framework could also make advanced materials research more accessible to smaller laboratories and institutions that do not have the computing resources available to large technology companies.“This will save huge computation costs and time by removing downstream selection requirements,” Mingda Li told MIT News. “That will help not only large efforts that generate hundreds of millions of materials, but also smaller research groups with targeted applications.”The system currently works best with highly ordered crystalline solids rather than every possible type of material. But the broader idea could prove significant: instead of using AI to generate everything and discard almost all of it later, researchers may be able to teach the system some of chemistry’s most important rules before it starts designing.Disclaimer: This article is based on research reported by MIT News and the study published in Nature Computational Science. The research findings and potential applications have not been independently verified by TOI Education.
