Source-linked AI summary
GraphAgents: Knowledge Graph-Guided Agentic AI for Cross-Domain Materials Design
Isabella A. Stewart, Tarjei Paule Hage, Yu-Chuan Hsu, Markus J. Buehler
TL;DR
Materials discovery is constrained by siloed expertise and limited cross-domain hypothesis generation, especially for difficult-to-replace PFAS. The paper combines specialized agents with knowledge-graph traversal, and finds that the complete multi-agent pipeline performs best overall in ablation studies while generating cross-domain design hypotheses.
Problem
Siloed expertise limits cross-disciplinary searches for viable substitutes, particularly for persistent PFAS whose property combination remains difficult to match.
Method
A multi-agent framework decomposes materials-design problems, retrieves and enriches evidence, extracts design parameters, and traverses knowledge graphs for hypothesis generation.
Results
The full multi-agent pipeline achieves the highest overall ablation performance, generally outperforming or matching configurations with components removed.
Takeaways & Limitations
Knowledge-graph-guided multi-agent reasoning supports cross-domain materials hypothesis generation by combining complementary retrieval, decomposition, evidence grounding, and relational exploration.
Takeaways & Limitations
The academic-corpus knowledge graphs propose scientifically novel composites that are not always industrially viable, motivating inclusion of patents and technical datasheets.
Abstract
from arXiv · showhide
Large Language Models (LLMs) promise to accelerate discovery by reasoning across the expanding scientific landscape. Yet, the challenge is no longer access to information but connecting it in meaningful, domain-spanning ways. In materials science, where innovation demands integrating concepts from molecular chemistry to mechanical performance, this is especially acute. Neither humans nor single-agent LLMs can fully contend with this torrent of information, with the latter often prone to hallucinations. To address this bottleneck, we introduce a multi-agent framework guided by large-scale knowledge graphs to find sustainable substitutes for per- and polyfluoroalkyl substances (PFAS)-chemicals currently under intense regulatory scrutiny. Agents in the framework specialize in problem decomposition, evidence retrieval, design parameter extraction, and graph traversal, uncovering latent connections across distinct knowledge pockets to support hypothesis generation. Ablation studies show that the full multi-agent pipeline outperforms single-shot prompting, underscoring the value of distributed specialization and relational reasoning. We demonstrate that by tailoring graph traversal strategies, the system alternates between exploitative searches focusing on domain-critical outcomes and exploratory searches surfacing emergent cross-connections. Illustrated through the exemplar of biomedical tubing, the framework generates sustainable PFAS-free alternatives that balance tribological performance, thermal stability, chemical resistance, and biocompatibility. This work establishes a framework combining knowledge graphs with multi-agent reasoning to expand the materials design space, showcasing several initial design candidates to demonstrate the approach.
1 Introduction
The introduction identifies siloed human expertise and PFAS persistence as barriers to developing sustainable material substitutes. It motivates combining knowledge graphs with coordinated, domain-specialized LLM agents to broaden materials exploration and generate testable hypotheses.
- Motivation: Narrow domain specialization limits scientists’ ability to envision replacements spanning polymers, ceramics, metals, biomaterials, and hybrid systems.Deep expertise within well-defined polymer design frameworks can come at the cost of cross-disciplinary breadth.
- Motivation: Siloed knowledge narrows materials-landscape searches and reduces chance encounters between disciplines, slowing hypothesis generation despite urgent demand for sustainable alternatives.The limitation arises from bounds on human expertise and imagination.
- Motivation: PFAS are difficult to replace because durability, chemical resistance, and water- and oil-repellency also contribute to environmental persistence and regulatory attention.PFAS are synthetic chemicals valued for these performance characteristics, while their persistence has prompted increasing regulatory attention in parts of the world.
- Approach: Ontological knowledge graphs represent concepts as nodes and relations as edges, extending LLM access to large-scale knowledge sources for scientific reasoning.The introduction presents knowledge graphs as a strategy for more efficient knowledge representation in generative AI for science.
- Approach: The proposed approach couples knowledge-graph idea mining with LLM generation and reasoning, enabling agents to traverse connections and uncover new relationships.The framework’s goal is to combine domain expertise with links across isolated knowledge pockets to broaden searches and produce creative, testable hypotheses.
2 Results and Discussion
The results show that knowledge-graph traversal supports multi-objective PFAS-free materials hypotheses, while different search strategies expose different levels of contextual and cross-functional information. Keyword matching remains imperfect, but the full multi-agent pipeline generally outperforms ablated configurations through distributed specialization and richer evidence integration.
- Multi-agent design analysis: The planner decomposes the multi-objective materials-design problem into intermediate subquestions and design parameters for downstream agents.This decomposition provides agents with background knowledge needed to address the larger design problem.
- Keyword matching and limitations: Graph matching links extracted requirements to related material-property nodes, but semantic mismatches can omit application-specific constraints.Examples include thermal stability mapped to temperature stability and biomedical durability mapped to Biological durability, losing sterilization and cyclic-temperature context.
- Graph traversal strategies: Shortest paths can yield hollow or self-referential edges, whereas Top-N Shortest Simple Paths (N = 5) provide richer context and broader cross-functional material proposals.The shortest-path proposal centered on PLA-CNF-PDA, while the Top-N proposal additionally integrated TPU, CNCs, and Al2O3 nanoparticles.
- Graph traversal strategies: DFS generates deeper cross-domain hypotheses, such as natural biopolymers reinforced with MWCNTs and functionalized with TiO2 nanoparticles for biomedical tubing.The extended paths connect biocompatibility, flexibility, protein adsorption, mechanical reinforcement, conductivity, microstructural stability, corrosion resistance, and biocompatibility.
- Ablation study: The full multi-agent pipeline achieves the highest overall performance, surpassing or matching ablated configurations across evaluation criteria except cross-subtask integration and novelty.Direct Graph Exploration exceeds the full configuration in novelty, while the full pipeline produces more grounded hypotheses from a broader retrieved knowledge base.
3 Conclusion and Future Directions
GraphAgents uses targeted in-context knowledge injections and PFAS knowledge graphs to guide multi-agent hypothesis generation and uncover latent cross-domain connections. Its current academic focus can limit industrial viability, motivating broader corpora, dynamically constructed reasoning graphs, and eventual integration with autonomous synthesis labs.
- Core framework: Targeted injections of structured knowledge, constraints, or examples shape LLM behavior and guide generated outputs for hypothesis generation.The pipeline synthesizes information around a design problem and extrapolates to adjacent knowledge domains.
- Core framework: PFAS knowledge graphs expose intrinsic material properties and enable agents to traverse broader knowledge spaces for previously unrecognized connections.In-context content is dynamically tuned using external knowledge bases and on-demand tools.
- Core framework: The complete five-agent configuration achieves the most balanced and robust performance, while removing components consistently reduces overall scores.The largest performance gap occurs between the full pipeline and the LLM Only setup; intermediate configurations show that individual agents contribute value.
- Limitations and future directions: Academic-corpus knowledge graphs favor scientifically novel composites such as graphene oxide, cellulose nanocrystals, and lignin blends that are not always industrially viable.Future work will add patents and technical datasheets containing industry-relevant testing protocols, performance standards, and material information.
- Limitations and future directions: Future iterations could let agents construct temporary reasoning graphs during inference to bridge distant domains only when necessary.The proposed direction includes in-situ knowledge expansion, such as Graph-PReFLexOR, and coupling the creative engine with autonomous synthesis labs.
4 Materials and Methods
The study builds two complementary knowledge graphs from PFAS literature: a depth-oriented PFAS-specific graph and a breadth-oriented Material Properties graph spanning 51 property categories. A pipeline converts literature into graph fragments, merges semantically similar entities, and equips agents with hybrid retrieval and heuristic graph traversal tools.
- Knowledge graph design: Two knowledge graphs provide complementary substrates: PFAS-specific literature supports targeted exploration, while the Material Properties graph enables cross-domain hypothesis generation across 51 property categories.The PFAS-specific graph emphasizes application rationale and performance requirements; the broader graph organizes knowledge around PFAS-relevant material properties.
- Graph construction: Scientific papers were converted to Markdown, segmented into context-limited sections, processed by an LLM to extract entity-relation triplets, and merged across papers into unified graph structures.Within each paper, correlated terminology and identical nodes connect extracted fragments; semantically similar nodes with cosine similarity greater than 0.95 are merged.
- Agent-tool integration: GraphWeave links document identifiers to connected subgraphs and maps input keywords to nearby nodes using embedding comparisons.Hybrid GraphWeave combines cosine-similarity retrieval from Chroma-stored text chunks with linked knowledge-graph subgraphs to ground agent context in textual evidence.
- Graph traversal: Heuristic traversal tools—including BFS with Semantic Stop, DFS, Shortest Simple Path, and Top-N Shortest Simple Path—vary the cross-domain context available for hypothesis generation.The implementation uses NetworkX for graph structures, Chroma for persistent embedding retrieval, and AutoGen for the multi-agent pipeline.
- Corpus construction: The corpora were assembled from Web of Science searches, yielding 4,824 PFAS-query results and 4,716 retrieved articles, corresponding to a yield of ≈97.8%.The broader graph used abstracts and metadata retrieved in batches of 1,000 records because full-text acquisition was impractical at its scale.