Chemically Meaningful Textualization Enables Explainable Validation of Metal-Organic Frameworks by Large Language Models

This paper demonstrates that large language models (LLMs) can serve as interpretable validators of metal-organic framework (MOF) structures when crystallographic information is transformed into chemically meaningful text. The authors show that unreasonable MOF structures can induce substantial deviations in computed energetic, electronic properties, and adsorption properties, highlighting the practical need for reliable validation. They benchmark nine descriptors spanning non-text, semi-text, and full-text formats and find that successful LLM-based validation depends not simply on the amount of structural information provided, but on whether that information is reorganized into a chemically meaningful and linguistically learnable form. Fine-tuned LLMs using specialized descriptors (mof2text) achieve performance comparable to graph-based models in identifying unreasonable MOFs. Importantly, these models extend beyond black-box classification by generating diagnostic rationales for likely error sources, including abnormal bonding, connectivity, and charge states, as well as error-category predictions for annotated datasets. The work establishes that chemically informed textualization provides a practical route to transforming general-purpose language models into explainable assistants for curating computation-ready MOF databases.

Improvements to AI systems:

1. Integrate domain-aware textualization modules into LLM pipelines for materials science, converting raw crystallographic data (CIF files) into structured, chemically meaningful text (e.g., bonding motifs, coordination environments, charge states) before validation tasks. This replaces naive tokenization of numeric coordinates.

2. Add an interpretable validation layer that outputs both a binary reasonable/unreasonable verdict and a natural-language diagnostic rationale (e.g., abnormal Zn–O bond length of 2.8 Å, missing hydrogen on carboxylate group). This turns black-box classifiers into explainable curators.

3. Implement error-category prediction heads on fine-tuned LLMs, enabling classification of specific failure modes (bonding, connectivity, charge, geometry) for each MOF, not just a pass/fail score.

4. Create a hybrid scoring system that combines LLM-based textual validation with graph-based models (e.g., GNNs) for cross-checking, using the LLM’s rationale to prioritize which structures need manual review or re-simulation.

5. Develop a feedback loop where the LLM’s generated rationales are used to automatically suggest corrections (e.g., reassign oxidation state of Cu from +2 to +1) or flag entries for database removal, improving the curation of computation-ready MOF databases.

What the improved AI system can do:

• Automatically screen thousands of hypothetical or experimental MOF structures from databases, flagging unreasonable ones with specific, human-readable reasons.

• Provide chemists with actionable diagnostics (e.g., unphysical bond angle at node, inconsistent metal coordination number) rather than a single confidence score.

• Reduce false positives in high-throughput computational screening by rejecting invalid structures before expensive DFT or adsorption simulations.

• Generate structured error reports for database curators, categorizing issues by type and frequency to guide synthesis or force-field parameter improvements.

• Serve as an explainable assistant that can answer follow-up questions like Why is this MOF invalid? and suggest targeted fixes, enabling non-experts to curate materials data reliably.

Computation-ready metal-organic framework (MOF) databases are essential for high-throughput screening, yet many reported crystal structures remain chemically unreasonable or disordered, compromising simulation fidelity. Existing validation approaches can identify non-computation-ready structures, but they often rely on heuristic rules, license requirement, or offer limited interpretability. Here, we show that large language models (LLMs) can serve as interpretable validators of MOF structures when crystallographic information is transformed into chemically meaningful text. By benchmarking nine descriptors, we find that successful LLM-based validation depends not on the amount of structural information alone, but on whether local coordination, framework connectivity, and chemical context are organized into a linguistically learnable representation. Fine-tuned LLMs using specialized descriptors (mof2text) achieve performance comparable to graph-based models in identifying unreasonable MOFs. Importantly, these models extend beyond black-box classification by generating diagnostic rationales for likely error sources, including abnormal bonding, connectivity, and charge states, as well as error-category predictions for annotated datasets. This work establishes chemically informed textualization as the key step that transforms LLMs from generic text models into practical and explainable tools for curating MOF databases.