Crystallography as language: transforming structural data into textual description for LLM-aided materials discovery

The discovery of high-performance permanent magnets necessitates effective strategies to stabilize promising compounds, such as SmFe12, which exhibit favorable magnetic properties but inherent thermodynamic instability. While data-driven methods have advanced materials discovery, they face significant limitations: combinatorial explosion of potential substitutions, descriptor design challenges, data scarcity and bias, and the opaque nature of many models lacking physical interpretability. This study presents a human-AI collaborative framework leveraging large language models (LLMs) to reason about crystal structures by transforming crystallographic data into interpretable textual descriptions. Our Prompt-Guided LLM approach achieves 85% classification accuracy and 88% F1-score across 96 SmFe12-based structures when validated against DFT calculations. The framework further demonstrates structure generation capability, producing 24 candidate structures, of which 10 exhibit enhanced stability or magnetization compared with their host structures; notably, SmFe12N and SmFe10Co2N – show enhancement in both properties compared with SmFe12 (I4/mmm). Crucially, the model provides physically sound rationales for its predictions, suggesting mechanisms such as lattice strain relief and electron balance optimization that align with established materials science principles. These results demonstrate that carefully implemented LLMs can effectively support crystal structure discovery while providing explanatory insights for materials design.