Imagine asking an artificial intelligence for a material to make better batteries and receiving a list of candidates.
Until now, finding substances with specific properties has required studying a multitude of chemical combinations and comparing their structures. The challenge is immense: small variations in the position of atoms can modify the stability, conductivity or resistance of a compound. And it is not enough to know its composition. How it is organized internally also matters.
MIT researchers have developed artificial intelligence that relates written descriptions to three-dimensional crystal structures, that is, how their atoms are arranged.
Their system learns to connect words about desirable characteristics with representations of real materials. In this way, a query expressed in common language can help locate structures compatible with certain requirements. It is a way of bringing together two very different ways of describing matter: through phrases and through atomic geometries.
When they searched for substances useful for solid-state batteries, the system identified options that other scientists already considered promising.
These batteries use a solid electrolyte, instead of the liquid usually used in many current batteries. Finding suitable materials is complicated because they must combine several properties at the same time: facilitate the movement of ions, maintain their stability and work together with the other components.
But recommending a compound does not mean that it can be manufactured or that it will work as we expect. That requires experiments.
Researchers must check how each candidate performs in real conditions. An attractive structure in a database could present difficulties in synthesis, degrade during use, or have insufficient performance. That is why artificial intelligence does not replace the laboratory: it helps decide what deserves to be studied first. This would save time and resources.
The key is to reverse the search: start with what we want to achieve and then find materials that could achieve it.
REFERENCES
Yang Jeong Park, Mayank Kumaran, Chia-Wei Hsu, Elsa Olivetti and Ju Li. “Scaling multimodal materials representation learning with synthetic narratives.” Nature CommunicationsAugust 26, 2026.
