Session: 06-17-05 AI Technology for Ocean Engineering - V

Submission Number: 181223

A Framework for Domain-Specific CAD/CAE Automation in Ship Design Powered by Large Language Models 

Research applying Large Language Models (LLMs) to Computer-Aided Design (CAD) automation is expanding, but current efforts primarily target general-purpose CAD systems. Significant challenges remain in specialized fields like ship design, where designers utilize domain-specific CAD platforms characterized by powerful but non-public commands and macro libraries. While these private assets are valuable to expert users, they impose steep learning curves that consume precious design time. Consequently, designers often expend considerable effort defining intended 3D geometry instead of focusing on the conceptual and creative aspects of design itself. This inefficiency ultimately reduces overall design productivity.

Therefore, this study proposes a new framework enabling natural language-driven 3D modeling within such specialized CAD environments and verifies its effectiveness. Specifically, it demonstrates how an LLM can effectively leverage a proprietary domain-specific knowledge base to generate valid C# code for ship structure modeling within NAPA Designer.

The proposed framework is architected upon three core components. First, it externalizes proprietary NAPA C# script documents into a searchable vector database, making the knowledge available to the LLM. Second, an LLM queries this database to extract relevant commands for synthesizing C# code based on the user's natural language prompt. Third, a prototype self-correction loop was implemented using GUI automation tools to read error messages from the NAPA Designer interface and feed them back to the LLM for iterative code refinement.

The effectiveness of this framework was validated through a series of modeling tasks. The developed system successfully generated valid C# code for the creation of ship structural components, demonstrating its ability to both take advantage of documented proprietary macros and generate ad hoc C# code in the absence of directly supported macros. This validation demonstrated that the LLM has the ability to interpret documented task-based commands and translate them into a series of valid and executable modeling operations.

This study demonstrates that a framework combining LLM, an external knowledge base, and a feedback loop can work effectively in a proprietary CAD environment. This approach opens new avenues for automation and efficiency in specialized ship design tasks by making proprietary commands available to the LLM. The validation of this framework is a proof of concept (PoC). It provides a concrete method to bridge the gap between modern AI capabilities and existing ship design platforms, and provides a path toward advanced AI-assisted design workflows in ship design.

Presenting Author: Kouki Wakita The University of Osaka

Presenting Author Biography: The presenting author, Dr. Kouki Wakita, is a full-time Specially Appointed Assistant Professor in the Open Collaboration Laboratory for Enabling Advanced Marine Systems (OCEANS) at the Graduate School of Engineering, the University of Osaka. He received his Ph.D. in Engineering from Osaka University in 2025 for his research on the maneuverability and autonomous control of ships. His research has been published in international journals such as "Journal of Marine Science and Technology" and "Ocean Engineering", and he has been actively involved in a JSPS-funded project on autonomous vessels. Dr. Wakita’s current research focuses on the automation of ship design using generative AI. By promoting next-generation research on ship design with advanced artificial intelligence technologies, he seeks to contribute to the advancement of environmentally sustainable maritime transportation and the future development of ocean engineering.

Authors:

Kouki Wakita The University of Osaka
Yasuo Ichinose The University of Osaka
Kosuke Hatayama The University of Osaka
Atsuo Maki The University of Osaka