GraphAgents: Knowledge Graph-Guided Agentic AI for Cross-Domain Materials Design

TL;DR

This work introduces GraphAgents, a knowledge-graph-guided multi-agent framework to address cross-domain materials design challenges, with a focus on identifying sustainable PFAS substitutes. By constructing PFAS-Specific and Material Properties knowledge graphs and coordinating planners, retrieval agents, evaluators, and hypothesis-generating graph-walkers, the method grounds design hypotheses in traceable evidence rather than opaque prompts. Ablation studies demonstrate that the full five-agent pipeline outperforms single-shot prompting, validating the value of distributed specialization and relational reasoning for multi-objective material design. The framework generates PFAS-free material hypotheses for biomedical tubing, balancing mechanical, thermal, chemical, transport, and biocompatibility requirements, and lays the groundwork for expanding knowledge-graph-driven discovery across domains and production pipelines.

Abstract

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.

Figures (17)

• Figure 1: Overview of the proposed agentic pipeline for materials discovery. The system begins with user input, which the Planner decomposes into sub-questions focusing on design parameters like material properties, operational conditions, failure modes, and interface requirements. The Hybrid GraphWeave agent retrieves evidence to answer each sub-question, from which the Evaluator selects pertinent design keywords with associated metrics. The Creative GraphWeave agent explores graph traversal pathways using the Evaluator keywords as stopping points to uncover novel connections and ideas. These are then synthesized by an Engineer into hypotheses that integrate both creative insights and contextual evidence. Bottom arrows illustrate how agents selectively retrieve and incorporate earlier responses as context for their own reasoning.
• Figure 2: Summary of the roles of each agent in the multi-agent system, from problem decomposition (Planner, HybridGraphWeave) to keyword extraction (Evaluator), idea generation (CreativeGraphWeave), and hypothesis formulation (Engineer).
• Figure 3: User–Planner Interaction. The Planner agent decomposes a broad user query on PFAS in biomedical tubing into sub-questions that guide the search for more precise answers.
• Figure 4: Complementary tool-calling workflows of the (A) Hybrid GraphWeave agent and (B) Creative GraphWeave agent. The Hybrid GraphWeave agent weaves evidence from two sources: the raw PFAS text corpus and the PFAS-Specific Knowledge Graph. It embeds the user query, retrieves the top matching text chunks from the corpus, and links them to corresponding subgraphs in the knowledge graph, thereby providing both textual evidence and relational context. In contrast, the Creative GraphWeave agent focuses on discovering new ideas by exploring and weaving connections between keywords within the knowledge graph. It identifies relevant nodes and applies pathfinding algorithms to assemble subgraphs that highlight potential connections. Together, the two agents provide complementary capabilities: Hybrid GraphWeave ensures evidence-grounded retrieval, while Creative GraphWeave enables exploratory ideation.
• Figure 5: Example of Evaluator agent output. The Evaluator agent selects and formats key material property descriptors with associated ranges or qualitative levels. These extracted keywords and metrics serve as design-relevant features that guide downstream agents in graph traversal and hypothesis generation.