Abstract
Pipe routing in shipbuilding and offshore engineering is a complex spatial planning task that must satisfy numerous functional requirements, including interference avoidance, maintainability, and operational accessibility. While algorithmic approaches like A* pathfinding frame this as a cost-minimization problem, determining appropriate penalty weights for factors such as path length, bend count, and deck proximity remains a highly subjective, labor-intensive process. Designers typically must adjust these parameters through continuous trial and error to meet project-specific constraints. To resolve this, this study introduces an automated pipe routing method integrated with a multi-agent LLM (Large Language Model) expert system. The proposed framework formalizes expert design knowledge—ranging from hard geometric constraints to soft layout preferences—into structured object and relation rules. A dedicated ‘result analysis agent’ qualitatively evaluates the algorithmic routing outcomes against these rules, while a ‘weight-tuning agent’ dynamically adjusts the A* cost weights based on the diagnostic feedback and historical iteration data. This closed-loop architecture enables the system to self-diagnose and iteratively re-optimize routes without human intervention. Verification tests demonstrate that the LLM accurately interprets spatial constraints and progressively refines routing quality. Over three iterations, the number of bends was reduced by approximately 14% and the grouping metric by approximately 73%, while the total path length increased by only about 8%, with each iteration requiring roughly one minute of computation. Additionally, the study confirms the practical feasibility of utilizing an on-premise open-source LLM to address industrial data security requirements. By successfully embedding expert reasoning directly into the optimization loop, this approach provides a highly adaptable, consistent, and automated framework for ship piping design.
