Leveraging Large Language Models for Natural Language-Driven CAD Automation and Multi-Objective Optimization in Thermal Component Design

The design of thermal components (such as automotive heat exchangers) requires balancing multiple competing objectives—thermal performance, aerodynamic efficiency, structural integrity, and manufacturability. Traditional design workflows rely on manual Computer Aided Design (CAD) modeling and iterative simulations, which are both labor-intensive and time-consuming. Recent advances in Large Language Models (LLMs) present untapped potential for automating parametric CAD generation. However, current LLM-based approaches primarily handle simple, isolated geometric primitives rather than complex multi-component assemblies. This work introduces a progressive framework that leverages fine-tuned LLMs (Qwen2.5-3B-SFT) integrated with the CadQuery CAD kernel to automatically generate parametric geometries from natural language descriptions. As a foundational study, this work focuses on Step 1 of the framework: generating and optimizing isolated geometric primitives (cylinders, pipes, etc.) that form the building blocks of complex assemblies. The generated models are automatically exported to standard CAD formats and seamlessly integrated within a multi-objective Bayesian optimization pipeline using Gaussian Process regression. By decoupling natural language-driven CAD code generation from traditional manual scripting, this work demonstrates how LLMs can accelerate design space exploration while eliminating the need for engineers to write geometry-specific Python scripts. Case studies on parametric pipe optimization demonstrate the framework's efficiency gains and establish a foundation for future steps: handling constraints, multi-component assemblies, and full thermal component designs. This work contributes to next-generation Artificial Intelligence (AI) assisted engineering design by demonstrating LLM-powered automation as a practical pathway toward fully automated design-to-optimization workflows.