Unlocking the secrets of crystal structures is a crucial step in predicting the properties and applications of materials from their chemical compositions. Unfortunately, most current methods for crystal structure prediction are highly computationally intensive, hindering the pace of technological advancement.
However, by seeding structure prediction algorithms with quality-generated candidates, we can effectively break through this major barrier.
Now, researchers at the University of Reading and University College London have introduced an innovative artificial intelligence model that can predict how atoms arrange themselves in crystal structures. Called CrystaLLM, the technology could lead to faster discovery of new materials for everything from solar panels to computer chips.
CrystaLLM operates similarly to AI chatbots, mastering the “language” of crystals through extensive analysis of millions of existing crystal structures. The researchers are committed to sharing this powerful tool with the scientific community, enabling faster and more efficient material discovery that could revolutionize technology as we know it.
“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,” said Dr Luis Antunes, who led the research while completing his PhD at the University of Reading. “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.”
The current process for determining how atoms arrange themselves into crystals is based on lengthy computer simulations that analyze the physical interactions of atoms.
However, CrystaLLM simplifies this process. By reading millions of crystal structure descriptions in Crystallographic Information Files—the industry standard for crystal representation—it circumvents complex physics calculations altogether.
CrystaLLM treats these structure descriptions as plain text. As it processes each description, it makes predictions about subsequent content, gradually identifying patterns in crystal formation. It was never instructed on any physics or chemistry principles but independently uncovered these concepts. It discovered how atoms are arranged and how their sizes influence the shape of crystals solely through reading these descriptions.
In testing, CrystaLLM successfully produced realistic crystal structures, even for substances it had not previously encountered. The research team has launched a free website allowing researchers to utilize CrystaLLM for crystal structure generation.
Incorporating this model into crystal structure prediction processes could accelerate the discovery of new materials for advancements such as improved batteries, more effective solar panels, and quicker computer processors.
Journal reference:
- Luis M. Antunes, Keith T. Butler & Ricardo Grau-Crespo. Crystal structure generation with autoregressive large language modeling. Nature Communications, 2024; DOI: 10.1038/s41467-024-54639-7