Research indexConference paper
Using Large Language Models for Compressor Performance Optimization
Wu, Zhang, Xu and 3 others
Volume 11: Turbomachinery — Multidisciplinary Design Approaches, Optimization, and Uncertainty Quantification; Radial Turbomachinery Aerodynamics15 Jun 2026Open access
Using Large Language Models for Compressor Performance Optimization
Wu, Zhang, Xu and 3 others · Volume 11: Turbomachinery — Multidisciplinary Design Approaches, Optimization, and Uncertainty Quantification; Radial Turbomachinery Aerodynamics · 2026
Abstract
Large language models (LLMs) exhibit remarkable in-context learning (ICL) capabilities, enabling them to flexibly leverage limited historical information and play a critical role in reasoning, problem solving, and complex pattern recognition tasks. These capabilities have led to growing interest in their potential for engineering optimization. Aerodynamic optimization is a key challenge in turbomachinery design. To evaluate the feasibility of applying LLMs to turbomachinery aerodynamic design, this study proposes an LLM-driven optimization framework that integrates prompt engineering with ICL strategies. This framework enables the LLM to autonomously generate, evaluate, and iteratively refine compressor geometric design parameters based on aerodynamic simulation feedback. Specifically, a two-dimensional subsonic cascade is selected as the benchmark case, and the MISES solver is employed for rapid performance evaluation. A structured prompt template is designed to dynamically provide key optimization information to the LLM, enabling it to efficiently generate high-quality design candidates that satisfy given constraints. Acting as a design generator, the LLM iteratively produces feasible solutions with progressively improved aerodynamic performance. Under identical configurations, multiple LLMs are systematically compared in terms of optimization stability, convergence speed, and solution quality. The results indicate that LLM-based optimization overall outperforms traditional PSO and GA algorithms, demonstrating stronger capability in understanding the design space and constructing efficient search strategies, thereby achieving superior optimization performance, with Grok, Gemini, and DeepSeek achieving the best results. These findings suggest that LLMs, as novel language-driven optimizers, offer a new methodological framework for intelligent aerodynamic design of turbomachinery and provide valuable insights for other geometry optimization problems.