Speaking crystal: AI learns language of atom arrangements in solids

The technology, called CrystaLLM, was developed by researchers at the University of Reading and University College London. It works similarly to an artificial intelligence chatbot, learning the “language” of crystals by studying millions of existing crystal structures.

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

Dr Luis Antunes, who led the research while completing his PhD at the University of Reading, said: “Predicting a crystal structure is like solving a complex, multi-dimensional puzzle where the parts are hidden. Crystal structure prediction requires a lot of computing power to test countless possible permutations.

“CrystaLLM achieves breakthroughs by studying millions of known crystal structures to understand patterns and predict new ones, much like a puzzle solver is able to identify winning patterns instead of trying every possible move.”

Predict the structure of unfamiliar materials

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

CrystaLLM treats these crystal descriptions as text. As it reads each description, it predicts what will happen next and gradually learns patterns about the crystal structure. The system was never taught any rules of physics or chemistry and instead calculated on its own. Just by reading these descriptions, it’s possible to understand how atoms are arranged and how their size affects the shape of the crystal.

When tested, CrystaLLM can successfully generate realistic crystal structures, even for materials that have never been seen before.

The research team has created a free website where researchers can use CrystaLLM to generate crystal structures. Integrating this model into crystal structure prediction workflows could accelerate the development of new materials to enable technologies such as better batteries, more efficient solar cells and faster computer chips.