Hypergraphs Turn LLMs into Reliable Material Discovery Agents

Motivation: Higher‑order relations in scientific knowledge

Traditional knowledge graphs model facts as head‑relation‑tail triples. Many scientific processes involve n‑ary interactions (e.g., a chemical reaction with 5‑6 reagents or a scaffold whose mechanical properties depend on PCL, chitosan, hydroxyapatite, porosity, and cell behavior). Converting such n‑ary facts into pairwise edges creates O(n²) edges, discards contextual information, and leads to hallucinations in retrieval‑augmented generation (RAG) pipelines.

Hypergraph representation

Each n‑ary relation extracted from a paper is stored as a hyperedge in a hypergraph H = (V, E). Hyperedges can connect any number of vertices, allowing downstream LLM agents to traverse the structure directly without losing context.

Data collection and preprocessing

Corpus : Web of Science query “biocomposite scaffold” → 1 297 full‑text PDFs.

Chunking : Each PDF split into non‑overlapping 10 k‑character blocks.

Extraction pipeline (LLaMA‑4‑17B‑128E):

Pass 1 – syntactic SVO parsing to produce explicit n‑tuples.

Pass 2 – semantic completion (nominalization → light verb, purpose phrases → “used‑for”).

Hypergraph construction and cleaning (Algorithm 1)

For each paper a local hypergraph Hᵢ is built; every ten papers the local graphs are merged into a global hypergraph H.

Embedding‑based deduplication (nomic‑embed‑text‑v1.5, cosine similarity ≥ 0.95) is performed on node labels. The term with the highest degree is kept as the canonical representative; synonyms (e.g., “PLA” ↔ “polylactic acid”) are merged.

Self‑loops and one‑way self‑references are removed while provenance information (source document, chunk) is retained.

Global hypergraph statistics

Number of nodes: ≈ 161 k

Number of hyperedges: ≈ 320 k

Average hyperedge size: 2.35 (maximum 32)

Degree distribution follows a power‑law with exponent γ≈1.23 (scale‑free corpus).

Pairwise co‑occurrence edges would explode to 22.1 M, but only 210 k triples appear three or more times, providing a natural noise filter.

Network insights

Top‑30 hub nodes correspond to core domain concepts such as “scaffolds”, “biocompatibility”, and “chitosan”.

s‑connected components (hyperedges sharing at least s vertices):

s = 1 → 153 k edges (largest component).

s = 4 → 4.5 k edges, automatically isolating mature formulation clusters.

Multi‑agent reasoning workflow (AutoGen + LLaMA‑3.3‑70B‑Q4)

Three specialized agents operate on the hypergraph:

GraphAgent : receives a query, extracts relevant keywords, maps them to nodes, and computes k‑shortest hyperpaths using Yen’s algorithm (intersection size S is configurable).

Engineer : translates each hyperpath into a mechanistic chain of operations.

Hypothesizer : generates a concrete experimental protocol that can be executed in the lab.

Case study 1 – Cerium oxide ↔ PCL

With s = 1 and K = 3 shortest hyperpaths, the common intermediate node is “chitosan”. The agents propose a PCL‑chitosan‑cerium‑oxide nanofiber scaffold and output a full electrospinning and antibacterial‑testing workflow.

Case study 2 – Grass ↔ PCL

The hypergraph reveals a pathway “fescue grass → hydrogen production → methanol → PCL precipitation solvent”. The Hypothesizer designs a closed‑loop green experiment using grass‑derived biomass, methanol as a green solvent, and PCL precipitation.

Scalability and generality

No domain‑specific fine‑tuning is required; topological constraints alone suppress LLM hallucinations.

The hypergraph can be incrementally updated as new papers are processed, keeping the knowledge base current.

The same pipeline can be applied to other scientific domains (e.g., batteries, proteins, alloys) by swapping the literature corpus.

Repository

https://github.com/lamm-mit/HyperGraphReasoning
Higher-Order Knowledge Representations for Agentic Scientific Reasoning
https://arxiv.org/pdf/2601.04878