What if finding the perfect material for a new battery was as easy as ordering dinner?
That’s the idea behind a new AI model from MIT. Most materials we rely on, from batteries to electronics, are crystals, which are atoms arranged in precise 3D patterns. The trouble is that today’s AI language models learn from text, not from 3D structures, and the scientific literature mostly covers a handful of well-studied crystal types. So the AI ends up with a narrow, flat view of the material world.
The MIT team built a model that can “see” both the atomic structure and the language describing it. They trained it to pair 3D crystal structures with rich written descriptions covering properties, how the material is made, and what it could be used for. It’s a bit like teaching a brain to link what it sees with what it reads.
The results are promising. When asked to find materials for solid-state lithium-ion battery electrolytes, the model’s top picks included materials that other researchers had independently flagged as promising. It also outperformed models trained on text alone.
The researchers hope it becomes for materials science what AlphaFold became for biology, a tool that connects structure to function and speeds up discovery.
If AI could design any material on demand, what would you want it to create first? 🔬
#Science #AI #MaterialsScience #MIT #Innovation
That’s the idea behind a new AI model from MIT. Most materials we rely on, from batteries to electronics, are crystals, which are atoms arranged in precise 3D patterns. The trouble is that today’s AI language models learn from text, not from 3D structures, and the scientific literature mostly covers a handful of well-studied crystal types. So the AI ends up with a narrow, flat view of the material world.
The MIT team built a model that can “see” both the atomic structure and the language describing it. They trained it to pair 3D crystal structures with rich written descriptions covering properties, how the material is made, and what it could be used for. It’s a bit like teaching a brain to link what it sees with what it reads.
The results are promising. When asked to find materials for solid-state lithium-ion battery electrolytes, the model’s top picks included materials that other researchers had independently flagged as promising. It also outperformed models trained on text alone.
The researchers hope it becomes for materials science what AlphaFold became for biology, a tool that connects structure to function and speeds up discovery.
If AI could design any material on demand, what would you want it to create first? 🔬
#Science #AI #MaterialsScience #MIT #Innovation
1
