CaMEL-RAG

Hackathon project — a retrieval-augmented LLM that answers natural-language questions about catalyst adsorption energies, grounded in Open Catalyst data.

Code: GitHub · Event: 2025 LLM Hackathon for Applications in Materials Science & Chemistry · Team: Code4Catalysis-KFUPM (Montassar Bouzidi, Nur Allif Fathurrahman, A. B. M. Ashikur Rahman, Tasnim Ahmed, Michail Mitsakis)

Catalyst-screening datasets such as Open Catalyst hold millions of DFT results, but a researcher can’t simply ask them a question. Our idea for the hackathon was a natural-language front end: ask about an adsorbate on a surface and get back a number traceable to the underlying calculation, rather than a fluent guess.

Workflow from the hackathon submission.

How it works

  • Knowledge base: records from an Open Catalyst-derived table (adsorption energies, descriptors and structure metadata) are embedded with all-MiniLM-L6-v2 sentence-transformer embeddings and indexed in FAISS.
  • Retrieval: each question retrieves its top-_k_ records. They are inserted into the prompt with [#doc_id] tags, so every answer cites the records it came from.
  • Generation: an LLM (GPT-4.1-mini by default; the model is swappable) answers using only the retrieved context.
500 test queries: the adsorption energies returned by CaMEL-RAG against the DFT values in the index.

Results, read carefully

On 500 test queries, the returned adsorption energies matched the DFT values exactly (R² = 1.00). These queries ask about records that are in the index, so the test shows that retrieval and grounding work: the model reports the stored number instead of making one up. It does not show that the model can predict energies for systems it hasn’t seen. The obvious next step is an evaluation on held-out and near-duplicate queries, and detecting out-of-distribution requests.

The hackathon’s outcomes, including this project, are summarised in the event paper on arXiv.

Stack: Python · FAISS · sentence-transformers · OpenAI API · pandas · Jupyter