Crystal structure generation with autoregressive large language modeling

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Predicting material properties from chemical composition requires accurate crystal structure generation.
• Current crystal structure prediction methods are computationally intensive, hindering rapid innovation.
• High-quality candidate structures are crucial for efficient structure prediction algorithms.

Purpose of the Study:

• To introduce CrystaLLM, a novel methodology for versatile crystal structure generation.
• To leverage large language modeling (LLM) for predicting crystal structures.
• To accelerate the discovery and innovation of new materials.

Main Methods:

• Developed CrystaLLM based on autoregressive large language modeling (LLM).
• Trained the model on millions of Crystallographic Information File (CIF) datasets.
• Modeled crystal structures as text sequences for LLM processing.

Main Results:

• CrystaLLM successfully generated plausible crystal structures for various inorganic compounds.
• The generated structures were validated using ab initio simulations.
• The methodology demonstrated effectiveness for compounds not seen during training.

Conclusions:

• CrystaLLM offers a computationally efficient approach to crystal structure generation.
• The study highlights the potential of LLMs in learning crystal chemistry effectively.
• This approach can significantly accelerate materials discovery and innovation.