Large language model agent as a mechanical designer

Traditional mechanical design relies on iterative refinement cycles driven by expert heuristics and computationally expensive Finite Element Method (FEM) analyses to meet performance targets. While machine learning approaches have been developed to automate portions of this process, they typically demand large datasets, substantial computational resources, and remain narrowly tailored to specific domains, thereby limiting their generalizability across diverse design problems. We propose a framework coupling a pretrained Large Language Model (LLM) with an FEM module to autonomously generate, evaluate, and refine structural designs based on performance specifications and quantitative feedback. Without domain-specific fine-tuning, the LLM leverages general reasoning to iteratively propose candidate structures, interpret simulation results, and implement mechanically informed modifications. Demonstrated on 2D truss structures, the approach navigates complex design spaces, balances competing objectives, and identifies appropriate stopping points when optimization yields diminishing returns. Compared to traditional methods like NSGA-II, LLM-guided optimization achieves faster convergence with fewer FEM evaluations in highly discrete, multi-faceted design spaces involving dynamic node generation and discrete member sizing. Experiments across multiple temperature settings and model scales reveal that smaller models achieve slightly better constraint satisfaction with fewer iterations, while lower temperatures produce more consistent performance. These findings position LLM agents as a promising new class of natural language-based, reasoning-driven optimizers for engineering design.