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

TL;DR

The paper tackles the challenge of coordinating cross-domain expertise in engineering design by proposing a knowledge-guided multi-agent framework that grounds reasoning in two specialized knowledge graphs. Three agents—Graph Ontologist, Design Engineer, and Systems Engineer—collaborate with a human Manager in a structured design loop to generate, review, and refine NACA airfoils, with tools like AeroSandbox and NeuralFoil supporting rapid analysis. Key contributions include constructing role-specific knowledge graphs from literature, implementing a RAG-based knowledge retrieval and multi-modal review process, and detailing a six-step workflow that integrates design-space sampling, optimization with Kulfan coordinates, and iterative revisions. The framework demonstrates end-to-end feasibility on a pedagogical airfoil-design problem, showing improvements in design quality, explainability, and human-in-the-loop governance, with potential to generalize to other engineering domains and design tasks.

Abstract

The engineering design process often demands expertise from multiple domains, leading to complex collaborations and iterative refinements. Traditional methods can be resource-intensive and prone to inefficiencies. To address this, we formalize the engineering design process through a multi-agent AI framework that integrates structured design and review loops. The framework introduces specialized knowledge-driven agents that collaborate to generate and refine design candidates. As an exemplar, we demonstrate its application to the aerodynamic optimization of 4-digit NACA airfoils. The framework consists of three key AI agents: a Graph Ontologist, a Design Engineer, and a Systems Engineer. The Graph Ontologist employs a Large Language Model (LLM) to construct two domain-specific knowledge graphs from airfoil design literature. The Systems Engineer, informed by a human manager, formulates technical requirements that guide design generation and evaluation. The Design Engineer leverages the design knowledge graph and computational tools to propose candidate airfoils meeting these requirements. The Systems Engineer reviews and provides feedback both qualitative and quantitative using its own knowledge graph, forming an iterative feedback loop until a design is validated by the manager. The final design is then optimized to maximize performance metrics such as the lift-to-drag ratio. Overall, this work demonstrates how collaborative AI agents equipped with structured knowledge representations can enhance efficiency, consistency, and quality in the engineering design process.

Figures (9)

• Figure 1: Schematic for LLM-based multi-agent workflow, consisting of three LLM agents namely Graph Ontologist, Design Engineer, Systems Engineer, each assigned a designated set of tasks in this process. The Design and Systems Engineer agents are provided with their respective knowledge graphs tailored specific to their roles. The Design Engineer agent possesses a set of engineering tools to help him accomplish his tasks. The Systems Engineer agent utilizes a multi-modal language model for reviewing airfoil shapes generated by the Design Engineer and provides design improvement suggestions. The Design Engineer then acts on the feedback, updates the design, and sends it for another round of review process. This process also allows for a Manager (human user in this case) to provide feedback for design improvement. The design-review cycle terminates once the Manager determines the design to be a suitable candidate for further optimization. Details for each component and steps in this workflow are discussed later in Section \ref{['sec:methodology']}.
• Figure 2: Prompts provided to Graph Ontologist for generating two different KGs for our engineering agents. For the Systems Engineer KG, a more generic set of goals leads to the creation of a larger domain of knowledge as compared to the Design Engineer's KG, which is more focused on aspects related to airfoil design.
• Figure 3: Representation of Knowledge Graphs generated by the Graph Ontologist for the two engineering agents. Systems Engineer's KG contains $\approx 700$ nodes while the Design Engineer's KG contains $\approx 50$ nodes. Note that these graphs only contain nodes with a degree of 10 or more, the remaining nodes were filtered out for compactness of the knowledge base.
• Figure 4: Initial airfoil samples generated by the Design Engineer agent using the NACA generator and plotter tool. The image shows 10 out of the 100 samples generated in this step following parameter selection in Stage 2.
• Figure 5: Initial filtered design samples selected for design review by the Systems Engineer agent.