黑瓜吃料网

A new artificial intelligence model that can predict how atoms arrange themselves in crystal structures could lead to faster discovery of new materials for everything from solar panels to computer chips.

The technology, called CrystaLLM, was developed by researchers at the 黑瓜吃料网 and University College London. It works similarly to AI chatbots, by learning the "language" of crystals by studying millions of existing crystal structures.

Published today (Friday, 6 December) in Nature Communications, the new system will be distributed to the scientific community to aid the discovery of new materials.

Dr Luis Antunes, who led the research while completing his PhD at the 黑瓜吃料网, said: “Predicting crystal structures is like solving a complex, multidimensional puzzle where the pieces are hidden. Crystal structure prediction requires massive computing power to test countless possible arrangements of atoms.

“CrystaLLM offers a breakthrough by studying millions of known crystal structures to understand patterns and predict new ones, much like an expert puzzle solver who recognises winning patterns rather than trying every possible move.”

Predicting structures for unfamiliar materials

The current process for figuring out how atoms will arrange themselves into crystals relies on time-consuming computer simulations of the physical interactions between the atoms. CrystaLLM works in a simpler way. Instead of using complex physics calculations, it learns by reading millions of crystal structure descriptions contained in Crystallographic Information Files - the standard format for representing crystal structures.

CrystaLLM treats these crystal descriptions just like text. As it reads each description, it predicts what comes next, gradually learning patterns about how crystals are structured. The system was never taught any physics or chemistry rules, but instead figured them out on its own. It learned things like how atoms arrange themselves and how their size affects the crystal's shape, just from reading these descriptions.

When tested, CrystaLLM could successfully generate realistic crystal structures, even for materials it had never seen before.

The research team has created a  where researchers can use CrystaLLM to generate crystal structures. The integration of this model within crystal structure prediction workflows could speed up the development of new materials for technologies like better batteries, more efficient solar cells, and faster computer chips.

Antunes, L.M., Butler, K.T. & Grau-Crespo, R. Crystal structure generation with autoregressive large language modeling. Nat Commun15, 10570 (2024).

Image credit: Ella Maru Studio