LLMs Validate Metal-Organic Frameworks via Chemical Textualization

A recent study published on arXiv (2608.11283) reveals that large language models (LLMs) can effectively validate metal-organic framework (MOF) structures when crystallographic data is translated into meaningful chemical text. This research tackles the challenge posed by computation-ready MOF databases that contain chemically implausible or disordered structures, which can undermine simulation accuracy. Current validation techniques often depend on heuristic rules, licensing, or lack interpretability. The authors evaluated nine descriptors, concluding that successful LLM validation hinges not just on structural information but also on the organization of local coordination, framework connectivity, and chemical context into a learnable linguistic format. Fine-tuned LLMs using specialized descriptors (mof2text) demonstrated performance on par with existing methods, providing a more transparent validation approach. This study emphasizes the promise of LLMs in materials science, especially for the validation of intricate crystalline structures.