Researchers at the University of Reading and University College London have developed a novel artificial intelligence model that can anticipate how atoms arrange themselves in crystal structures, which was published in Nature Communications.
Image Credit: everything possible/Shutterstock.com
The new technology, called CrystaLLM, could speed up the identification of new materials for various applications, including solar panels and computer chips.
It functions similarly to AI chatbots, learning the “language” of crystals by studying millions of existing crystal formations.
The new method will be offered to the scientific community to aid in developing new materials.
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 recognizes winning patterns rather than trying every possible move.
Dr. Luis Antunes, Department of Chemistry, University of Reading
Predicting Structures for Unfamiliar Materials
The current method for determining how atoms will arrange themselves into crystals is based on time-consuming computer simulations of physical interactions between the atoms. CrystaLLM operates in a simpler manner. Instead of performing sophisticated physics calculations, it learns by reading millions of crystal structure descriptions stored in Crystallographic Information Files, the industry-standard format for expressing crystal structures.
CrystaLLM treats these crystal descriptions exactly like text. As it reads each description, it anticipates what will happen next, gradually learning patterns about how crystals are built. The system was never taught any physics or chemistry rules; instead, it found them out for itself. It learned about how atoms arrange themselves and how their size impacts the structure of the crystal simply by reading these descriptions.
When tested, CrystaLLM could build realistic crystal structures from hitherto unknown materials.
The research team has developed a free website where researchers can utilize CrystaLLM to produce crystal structures. Incorporating this model into crystal structure prediction workflows could accelerate the development of new materials for technologies such as improved batteries, more efficient solar cells, and faster computer chips.
Journal Reference:
Antunes, L. M., et. al. (2024) Crystal structure generation with autoregressive large language modeling. Nature Communications. doi.org/10.1038/s41467-024-54639-7