MIT AI model suggests recipes for novel materials

14:50

By: Dakir Madiha

MIT AI model suggests recipes for novel materials

Researchers at the Massachusetts Institute of Technology have unveiled DiffSyn, an AI model that proposes promising synthesis pathways for complex materials, tackling one of the most time-intensive bottlenecks in materials science. Detailed in a study published today in Nature Computational Science, the tool achieved top accuracy in predicting creation routes for zeolites, materials key to catalysis, adsorption, and ion exchange processes.

"For an analogy, we know what kind of cake we want to bake, but right now we don't know how to do it," said lead author Elton Pan, a PhD candidate in MIT's Department of Materials Science and Engineering. "Material synthesis today relies on domain expertise and trial-and-error."

While companies like Google and Meta have leveraged generative AI to generate vast databases of theoretical materials with desirable properties, turning those into reality often demands weeks or months of painstaking lab work. DiffSyn speeds this up by drawing on over 23,000 synthesis recipes from 50 years of scientific literature.

The model employs a diffusion-based approach akin to DALL-E's image generation. Scientists input a target material structure, and DiffSyn outputs viable combinations of reaction temperatures, durations, precursor ratios, and other parameters. It can generate 1,000 synthesis pathways in under a minute, far outpacing traditional case-by-case methods.

To validate it, the team synthesized a new zeolite-type material using DiffSyn's suggestions. Tests showed enhanced thermal stability and promising morphology for catalytic applications.

Unlike prior machine learning models that linked materials to single recipes, DiffSyn accounts for multiple viable paths to the same structure. "This is a paradigm shift from one-to-one structure-synthesis matching to one-to-many," Pan explained. "That's a key reason for our strong benchmark gains."

The work received support from MIT International Science and Technology Initiatives, the National Science Foundation, ExxonMobil, and Singapore's Agency for Science, Technology and Research.

Looking ahead, the team sees potential expansion beyond zeolites to metal-organic frameworks and inorganic solids. "Ultimately, we'd connect these smart systems to real-world autonomous experiments, using agentic reasoning on experimental feedback to dramatically accelerate materials design," Pan said.