LitMOF: An LLM Multi-Agent for Literature-Validated Metal-Organic Frameworks Database Correction and Expansion

LitMOF is a large language model-driven multi-agent framework that validates and corrects structural errors in metal-organic framework databases by cross-referencing original literature, resulting in a curated dataset of nearly 187,000 computation-ready structures and the discovery of over 12,000 previously unrecorded experimental MOFs, thereby preventing systematic errors in materials discovery workflows.

The Gist

Imagine you are trying to build a massive, perfect library of tiny, sponge-like structures called Metal-Organic Frameworks (MOFs). Scientists use these sponges to catch pollution, store energy, or clean water. To find the best ones, they use super-fast computers to simulate how they work.

But here's the problem: The library is messy.

Over the years, scientists have deposited thousands of these sponge designs into databases. However, a recent check revealed that nearly half of them are broken. Some have missing atoms, some have atoms glued together in impossible ways, and others are just copies of the same sponge with the wrong number of pieces. It's like trying to build a house using blueprints where some walls are floating in mid-air and others are made of invisible ink.

If you try to run a computer simulation on a broken blueprint, the results are garbage. You might think a sponge is amazing at catching CO2, when in reality, it's a dud. This has been slowing down scientific discovery for years.

Enter LitMOF: The "AI Librarian" Team

The authors of this paper created a solution called LitMOF. Think of LitMOF not as a single robot, but as a team of specialized AI detectives working together, led by a project manager.

Here is how this team works, using a simple analogy:

1. The Team Members (The Agents)

Instead of one AI trying to do everything, they split the work:

• The Database Reader: This agent is the "Archivist." It goes to the official library (the CSD database) and pulls out the original blueprints (CIF files) for a specific sponge.
• The Paper Reader: This agent is the "Investigator." It finds the original scientific article where the sponge was invented. It reads the text, looking for clues about what the sponge should look like. It's like reading the architect's notes to see what they intended to build.
• The Reference Builder: This agent is the "Blueprint Designer." It takes the notes from the Investigator and the Archivist to draw a perfect, ideal version of the sponge. This is the "Gold Standard."
• The Inspector & Editor: This is the "Quality Control Inspector." It compares the messy, broken blueprint from the library against the perfect "Gold Standard."
  • If a wall is missing? It adds it.
  • If two atoms are too close and crashing? It moves them apart.
  • If the blueprint has a messy scribble (disorder)? It tries to figure out the most logical way to fix it.
• The Simulation Runner: Once the blueprint is fixed, this agent can immediately test it in a virtual wind tunnel to see how well it works.

2. The Magic Trick: "Plan-and-Execute"

How do they talk to each other? They use a method called "Plan-and-Execute."
Imagine you ask a human assistant: "Fix this broken house."
A normal computer might just guess. But LitMOF's team leader says: "Okay, let's make a plan. First, check the archives. Second, read the architect's notes. Third, compare them. Fourth, fix the errors."
If the first step fails (e.g., the notes are missing), the team leader doesn't panic. It says, "Okay, Plan B: Let's look for similar houses to guess what the missing notes said." This flexibility allows them to fix things that old, rigid computer programs couldn't touch.

What Did They Achieve?

By using this AI team, they did three huge things:

1. They Fixed the Broken Library: They took the existing database of experimental sponges and fixed 8,771 broken entries. These were previously "unusable" for computers. Now, they are perfect, ready-to-use models.
2. They Found Hidden Treasures: They discovered 12,646 new sponges that scientists had written about in papers but never actually uploaded to the database. It's like finding a secret room in the library full of blueprints that nobody knew existed.
3. They Proved It Matters: They tested this on a real-world problem: Direct Air Capture (sucking CO2 out of the sky).
   • When they used the broken blueprints, the computer thought some sponges were terrible and others were miracles.
   • When they used the fixed blueprints, the results changed completely. The "miracles" were actually duds, and some "duds" turned out to be the best candidates.
   • The Lesson: If you don't fix the data, you waste years of research chasing false leads.

The Big Picture

This paper isn't just about fixing a database; it's about a new way of doing science. Instead of humans manually checking thousands of files (which is impossible), we now have an AI team that reads the original literature, understands the context, and repairs the data automatically.

It turns a messy, error-prone library into a clean, reliable foundation for the next generation of materials science. It's the difference between trying to build a rocket with a crumpled, coffee-stained map versus having a GPS that updates itself in real-time.

Technical Summary

1. Problem Statement

Metal-Organic Frameworks (MOFs) are critical porous materials for applications like gas storage and separation. Their discovery relies heavily on large, curated databases (e.g., CoRE MOF, CSD MOF Subset) used for high-throughput screening and machine learning. However, recent analyses reveal a critical flaw: nearly half of the entries in these databases contain substantial structural errors.

• The Scale of Error: A study by White et al. found that 51% of entries across 14 major MOF databases violate basic chemical valence principles.
• Limitations of Current Methods: Existing curation pipelines (e.g., rule-based sanity checks, MOSAEC) are designed to identify and discard invalid structures but lack the capability to repair them. They rely on fixed heuristics and cannot reconcile discrepancies between crystallographic files (CIFs) and the original scientific literature.
• Consequence: This leads to the exclusion of valid experimental materials and the inclusion of chemically impossible structures, causing systematic errors in property prediction (e.g., adsorption energies) and misleading high-throughput screening results.

2. Methodology: The LitMOF Framework

The authors introduce LitMOF, the first Large Language Model (LLM) driven multi-agent framework capable of automatically detecting, validating, and repairing structural errors in MOF databases by cross-referencing primary literature.

Architecture

LitMOF employs a Plan-and-Execute architecture orchestrated by a Supervisor agent, which coordinates five specialized agents:

1. Database Reader: Retrieves metadata (DOI, Refcode, formula) from the Cambridge Structural Database (CSD), CoRE MOF, and MOSAEC-DB.
2. Paper Reader: Extracts structural information from the original publication associated with the MOF.
   • Innovation: Instead of Retrieval-Augmented Generation (RAG), it uses full-document inference on parsed text (PDF/HTML/XML) to capture global structural context, which is crucial for complex chemical descriptions.
   • Dynamic Prompting: Iteratively refines extraction if initial results are incomplete or mismatched with database records.
3. Reference Builder: Constructs a "Reference Graph" representing the ideal, chemically valid minimal repeating unit of the MOF. It synthesizes data from the Paper Reader (structural formulas, expanded abbreviations) and Database Reader, converting chemical names to graph objects using PubChem APIs and IUPAC parsers.
4. Inspector & Editor: Compares the CIF graph against the Reference Graph to identify and correct three specific error types:
   • Bond Errors: Adjusts distance thresholds to correct missing or extra bonds.
   • Hydrogen Errors: Corrects misplaced or missing hydrogen atoms using identity mapping and graph matching.
   • Disorder Errors: Resolves unresolved disorder (duplicated fragments, fractional occupancies) by enumerating candidate configurations and selecting the lowest-energy structure using Machine Learning Interatomic Potentials (MLIP).
5. Simulation Runner: (Optional) Executes downstream computational tasks (e.g., DFT geometry optimization, pore analysis) on the corrected structures.

Workflow

When a user queries a MOF (e.g., "Fix MOF PICLAS"), the Supervisor generates a plan:

1. Retrieve CSD records and the associated paper.
2. Extract structural data and match it to the CSD Refcode.
3. Build a reference graph.
4. Detect discrepancies (e.g., stoichiometry mismatches, disorder) and apply corrections.
5. Return a validated, computation-ready CIF.

3. Key Contributions

• LitMOF-DB: The construction of a curated database of 186,773 computation-ready experimental MOFs.
  • Derived from the CSD MOF Subset (128,799 entries).
  • Includes 8,771 previously invalid CoRE MOF entries that were successfully repaired (accounting for 65.3% of the "Not-Computation-Ready" CoRE entries).
  • Provides multiple variants: Free-Solvent-Removed (FSR), Bound-Solvent-Removed (BSR), and Ion-restored (ION).
• Discovery of Missing MOFs: The system identified 12,646 experimentally reported MOFs that were synthesized and characterized in literature but never deposited as CIF files in the CSD. It reconstructs these by identifying "parent" MOFs and the chemical transformations (e.g., linker exchange, metal substitution) required to generate the missing structures.
• Scalable Repair Paradigm: Demonstrates that LLM-driven agents can automate the reconciliation of unstructured text (papers) with structured data (CIFs), a task previously impossible at scale due to manual effort requirements.

4. Results and Validation

• Correction Success Rates:
  • Bond Errors: 87.6% success rate (2,291/2,616).
  • Hydrogen Errors: 96.8% success rate (21,235/21,932).
  • Disorder Errors: 18.9% success rate (2,177/11,508). Note: Disorder is inherently difficult as it often requires breaking experimental constraints to define a unique topology.
• Manual Validation: A random sample of 500 corrected MOFs showed a 98.2% success rate upon manual review against original publications.
• Impact on Screening (Direct Air Capture Case Study):
  • The authors screened MOFs for Direct Air Capture (DAC) of CO2.
  • Original (Uncorrected) Data: Showed severe overestimation of adsorption heat ($|Q_{st}|$), with 97 MOFs yielding infinite values due to structural artifacts.
  • Corrected Data: Produced physically reasonable adsorption energies.
  • Ranking Distortion: The Pearson correlation between original and corrected adsorption energies was only 0.056.
  • False Positives/Negatives: Screening on uncorrected data resulted in 160 missing candidates and 56 false positives compared to the corrected dataset. This proves that structural errors fundamentally distort material ranking and lead to the omission of high-performance candidates.

5. Significance

• Paradigm Shift in Curation: Moves the field from "discard-based" curation (filtering out bad data) to "repair-based" curation (fixing bad data), thereby expanding the accessible chemical space.
• Reliability of Data-Driven Discovery: Establishes that structural fidelity is a prerequisite for reliable machine learning and high-throughput screening. Using uncorrected databases leads to systematic misranking and wasted computational resources.
• Generalizability: The LitMOF framework provides a blueprint for self-correcting scientific databases across diverse materials classes, integrating structured repositories with unstructured scientific literature to create dynamic, evolving knowledge bases.

In conclusion, LitMOF resolves long-standing issues of structural fidelity in MOF databases, providing a scalable pathway to transform raw experimental data into reliable, computation-ready resources for accelerated materials discovery.