Toward Autonomous Engineering Design: A Knowledge-Guided Multi-Agent Framework¶

🕒 Published (v1): 2025-11-05 04:55 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

A three-agent LLM framework formalizes iterative engineering design via a structured Generator–Critic loop: a Graph Ontologist builds domain knowledge graphs from literature, a Design Engineer proposes candidates using those graphs and simulation tools, and a Systems Engineer reviews against requirements until a human Manager validates the design. The approach is demonstrated on NACA 4-digit airfoil optimization for maximum lift-to-drag ratio.

Problem¶

LLMs lack the deep domain knowledge and tool integration required for engineering design, while classical multi-agent systems (MAS) require prohibitively hand-crafted knowledge bases. The gap is: how to equip LLM agents with structured, literature-grounded expertise and connect them to validated simulation tools within an iterative design-review cycle—without relying solely on LLM parametric memory, which is prone to hallucination.

Method¶

Three specialized LLM agents operate in a fixed orchestration pipeline:

Graph Ontologist — Processes 50 NACA airfoil papers segmented into 3000-token chunks, extracting knowledge graph (KG) triples of the form {node1, node2, edge} representing semantic relationships. Two role-differentiated KGs are produced via distinct prompts: a broad KG for the Systems Engineer (covering design lifecycle, fabrication, validation) and a focused KG for the Design Engineer (design variables → performance metrics).
Design Engineer — Given its KG and callable tools (AeroSandBox for CFD analysis, NeuralFoil for ML-based aerodynamic surrogate), proposes candidate airfoils parameterized by the design space \(\mathcal{D} := [0.01, 0.095] \times [0.05, 0.9] \times [0.01, 0.40]\) for \((c_\text{max}, c_\text{loc}, t_\text{max})\).
Systems Engineer — Translates the Manager's natural-language requirements into formal technical specs, then reviews each candidate using its broader KG plus a multimodal LLM for visual shape inspection, returning qualitative and quantitative feedback.

The Design Engineer–Systems Engineer loop iterates until the Manager approves a valid candidate, which then enters a downstream optimizer for performance tuning.

Key Contributions¶

Automated ontology extraction pipeline (Graph Ontologist) that converts domain literature into structured KG triples for agent grounding
Role-differentiated KGs generated from the same corpus via distinct prompts, encoding different expertise levels (generalist vs. specialist)
Integration of external simulation tools (AeroSandBox, NeuralFoil) as agent-callable functions within the iterative loop
Human-in-loop Manager checkpoint as the explicit termination condition, separating conceptual design from optimization phases
Instantiation of the Generator–Critic MAS pattern in a quantitative engineering optimization context with formal requirements

Results¶

The paper is a framework/demonstration contribution; the provided text does not report comparative benchmark numbers. The framework is applied to NACA 4-digit airfoil design targeting maximum lift-to-drag ratio at Mach 0.6 and \(Re = 5 \times 10^6\) at 0° angle of attack. No ablation or baseline comparisons appear in the provided excerpt.

Limitations¶

Restricted to one low-dimensional (3-parameter), well-studied problem; generalization to other engineering domains is not empirically validated
KGs are static artifacts built once from 50 papers; no mechanism for incremental updates as literature grows
Termination requires a human Manager—not fully autonomous
No ablation quantifying KG contribution vs. KG-free LLM agents
Assumes 0° angle of attack, limiting aerodynamic realism
Scalability to high-dimensional design spaces (e.g., full 3D geometry) is unaddressed

Relevance to Harnesses / Meta-Harnesses¶

This framework is a concrete domain-specific multi-agent harness: a fixed orchestration pipeline where specialized agents hand off structured artifacts (KG triples → design candidates → review feedback) through an iterative loop with a human-controlled termination gate. The Graph Ontologist stage is directly analogous to a preprocessing harness that transforms raw documents into structured knowledge for downstream agents. The design-review feedback loop instantiates the Generator–Critic harness pattern with quantitative exit criteria, making it a useful reference design for anyone building iterative refinement harnesses over domain tools with human-in-loop checkpoints.