Research map / Graph learning / Knowledge and temporal graphs
Construction and Application of Materials Knowledge Graph in Multidisciplinary Materials Science via Large Language Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds a materials knowledge graph (MKG) of about 162,605 nodes and 731,772 edges by fine-tuning LLMs (Darwin, LLaMA, LLaMA2) for joint NER/RE on energy-materials abstracts, plus entity resolution and Jaccard/TransE-based link prediction.
Problem
Materials science knowledge is scattered across millions of papers, and existing material KGs rely on predicted relations that may not reflect real instances, and are hard to update, while computational databases often disagree with experiments.
Approach
An LLM fine-tuned on minimally annotated abstracts performs NER and RE jointly, emitting DOI-traceable triples; these are normalized with ChemDataExtractor, mat2vec, and an expert dictionary, then used for enhanced-Jaccard and TransE link prediction of material-application pairs.
Key results
- Darwin fine-tuned model reaches NER F1 0.7974 and RE F1 0.7002 versus LLaMA2 7b at 0.7541 and 0.6557; with ER normalization Darwin (ER) reaches NER F1 0.9296 and RE F1 0.8827.
- MKG built with 162,605 nodes and 731,772 edges from 150,000 abstracts.
- Using 2014 data, 48.5% of top-200 material-application predictions were validated within nine years.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply ER and normalization to Darwin | Darwin · NER evaluation | F1 score: 0.7974 -> 0.9296 | Helped |
| Apply ER and normalization to Darwin | Darwin · RE evaluation | F1 score: 0.7002 -> 0.8827 | Helped |
| Apply ER and normalization to Darwin | Darwin · ER evaluation | F1 score: 0.4571 -> 0.8913 | Helped |
| Use Darwin rather than Llama2 7b | Fine-tuned LLM · NER evaluation | F1 score: 0.7541 vs. 0.7974 | Helped |
| Remove ER-N/F (equivalent to adding ER-N/F) (ablation) | Darwin (Normalized) · NER evaluation | F1: 92.96 -> 92.96 | No clear effect |
| Add ER-N/F (ablation) | Darwin (Normalized) · RE evaluation | F1: 83.29 -> 88.27 | Helped |
| Remove ER-NF/A (equivalent to adding ER-NF/A) (ablation) | Darwin (Normalized) · NER evaluation | F1: 84.01 -> 92.96 | Helped |
| Add ER-NF/A (ablation) | Darwin (Normalized) · RE evaluation | F1: 83.07 -> 88.27 | Helped |
| Add ER-NF/A (ablation) | Darwin (Normalized) · ER evaluation | F1: 84.54 -> 89.13 | Helped |
| Add ER-ED expert dictionary (ablation) | Darwin (Normalized) · NER evaluation | F1: 88.59 -> 92.96 | Helped |
| Add ER-ED expert dictionary (ablation) | Darwin (Normalized) · RE evaluation | F1: 80.20 -> 88.27 | Helped |
| Add ER-ED expert dictionary (ablation) | Darwin (Normalized) · ER evaluation | F1: 50.00 -> 89.13 | Helped |
Surprising findings
- Plain LLM extraction performs weakly on entity resolution (ER F1 about 0.46), which the authors attribute to limited contextual memory and motivate extra ER steps.
Limitations
- Only abstracts were used as input, omitting experimental methods and data.
- Human validation covered only 500 randomly sampled triples, and entity/relation error rates vary widely across labels (e.g., Property and Descriptor have the highest disagreement).
- Evaluation of the fine-tuned LLM used only 60 annotated abstracts.
Experimental setup
Datasets: Web of Science energy-materials abstracts (150,000) expert-annotated abstracts (75) expert dictionary (~600 terms)
Benchmarks: 60 annotated abstracts (NER, RE, ER) MKG link-prediction validation against later publications
Metrics: precision recall F1 score percentage of predicted material-application pairs reported
Baselines: MatBERT LLaMA 7b LLaMA2 7b MatKG2 Jaccard similarity TransE
Builds on: ChemDataExtractor mat2vec TransE DBSCAN LLaMA
Closest papers by meaning
- MatExpert: Decomposing Materials Discovery By Mimicking Human Experts (ICLR 2025 · similarity 0.86)
- Zero-Shot Learning for Materials Science Texts: Leveraging Duck Typing Principles (AAAI 2025 · similarity 0.84)
- Periodic Materials Generation using Text-Guided Joint Diffusion Model (ICLR 2025 · similarity 0.84)
- Code-Generated Graph Representations Using Multiple LLM Agents for Material Properties Prediction (ICML 2025 · similarity 0.84)
- FlowLLM: Flow Matching for Material Generation with Large Language Models as Base Distributions (NeurIPS 2024 · similarity 0.84)
- Generative Hierarchical Materials Search (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.