Demonstrates improved code generation in engineering design using LLMs and agent collaboration in Modelica models, suggesting a novel approach to optimization.
Key Points
The aim is to enhance 1D CAE modeling through the use of large language models and agent collaboration.
Developed a framework using Retrieval-Augmented Generation to search Modelica repositories.
Utilized three collaborative agents: planner, coder, and tester, to automate library assembly.
Conducted experiments with OpenAI-4o to evaluate code generation performance with proprietary libraries.
Agent-based workflow outperformed RAG in generating candidate code for complex tasks.
Code generation performance improved significantly when using internal libraries.