Mini Materials Knowledge Graph (LLM-assisted Ontology)

A short, beginner-friendly project to dive into semantic modeling for materials KG.

🎯 Goal

Build a tiny, end-to-end pipeline that:

1. defines a small ontology for a narrow materials sub-domain,
2. ingests a tiny dataset (abstracts or tabular snippets),
3. uses an LLM-assisted extractor to propose entities/relations,
4. converts outputs into RDF triples (Linked Data),
5. loads into a lightweight triplestore and runs a few SPARQL queries.

Target skill: ontology design, semantic consistency, FAIR/linked data, and LLM support—

🧭 Roadmap (minimal scope)

• Ontology: start with 6–10 classes (e.g., Material, SynthesisMethod, Property, Publication) and 10–20 relations (e.g., hasProperty, synthesizedBy, measuredIn).
• Sample data: a handful of abstracts or 1–2 CSV tables.
• LLM-assisted parsing: prototype prompts/rules to extract (subject, predicate, object) candidates.
• RDF conversion: map to a namespace, produce .ttl or .rdf.
• Queries: run 3–5 SPARQL queries that demonstrate retrieval & simple reasoning (e.g., “materials synthesized via solvothermal with band gap > X”).
• (Optional) Basic visualization (screenshot or simple graph view).

Keep it tiny and clean. Depth over breadth.

🛠️ Stack (flexible)

• Python: rdflib for RDF generation.
• SPARQL: rdflib SPARQL in-notebook or a local triplestore (e.g., GraphDB Free, QLever Docker, or Apache Jena/Fuseki).
• LLM parsing: Chatgpt and OLLAMA tried. Ollama is used, chatgpt does not provide api access for a plus user as in my case.
• Validation: simple checks for semantic consistency (domain/range sanity).

📦 Minimal Setup

# after cloning
python -m venv .venv && source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install rdflib pandas
# (Optional for local store) GraphDB/Jena/QLever via Docker if desired

🔍 Planned Sub-domain Ideas

• MOFs 101 (very small set)
• Common semiconductors (Si, GaAs, perovskites)
• Alloys (Fe-based) with 2–3 properties (e.g., density, hardness, band gap if relevant)

🧪 Notebook

All code will live in: notebooks/build_mini_mkg.ipynb

📚 Useful References (for later reading)

• MatVis (BAM-linked): https://github.com/Mat-O-Lab/MatVis
• Propnet (Materials KG/inference): https://github.com/materialsintelligence/propnet
• MatKG (materials KG from literature): https://www.nature.com/articles/s41597-024-03039-z and https://arxiv.org/abs/2210.17340
• MKG via LLMs (2024): https://arxiv.org/abs/2404.03080
• KG-FM (framework materials KG): https://www.nature.com/articles/s41524-025-01540-6
• Semi-automated KG pipeline (2025): https://pubs.rsc.org/en/content/articlehtml/2025/dd/d4dd00362d

Author: Md. Saidul Islam