• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (18): 84-98.doi: 10.3901/JME.260956

• 特邀专栏:大模型驱动的高端装备智能运维 • 上一篇    

扫码分享

大语言模型赋能优化设计初探

王沐晨1,2, 朱泓宇1,2, 宋学官1,2   

  1. 1. 大连理工大学机械工程学院 大连 116024;
    2. 高性能精密制造全国重点实验室 大连 116024
  • 收稿日期:2025-10-25 修回日期:2026-06-16 发布日期:2026-11-20
  • 作者简介:王沐晨,女,1995年出生,博士研究生。主要研究方向为基于深度学习的物理场信息重构与预测。E-mail:wangmuchen@mail.dlut.edu.cn;朱泓宇,男,1997年出生,博士研究生。主要研究方向为深度学习与形性一体化数字孪生。E-mail:hongyuzhu@mail.dlut.edu.cn;宋学官(通信作者),男,1982年出生,博士,教授,博士研究生导师。主要研究方向为重大装备及基础件的性能建模、优化设计、人工智能与装备数字孪生。E-mail:sxg@dlut.edu.cn
  • 基金资助:
    国家重点研发计划(2023YFB3406900)和国家杰出青年科学基金(52425509)资助项目。

Initial Exploration of Large Language Models-Enabled Optimization Design

WANG Muchen1,2, ZHU Hongyu1,2, SONG Xueguan1,2   

  1. 1. School of Mechanical Engineering, Dalian University of Technology, Dalian 116024;
    2. State Key Laboratory of High-performance Precision Manufacturing, Dalian 116024
  • Received:2025-10-25 Revised:2026-06-16 Published:2026-11-20

摘要: 大语言模型(Large language model,LLM)凭借其强大的自然语言处理能力、知识整合以及生成能力,正在迅速改变多个领域的研究范式与实践方式。然而,其在工程分析与优化设计中的应用潜力尚未得到充分探索。聚焦于LLM与优化设计的结合,通过常规的试验设计,包括9个标准测试函数、12项优化任务以及2个工程案例,对DeepSeek、ChatGPT、豆包、Gemini、Grok、Kimi六种主流LLM在预测与优化任务中的性能进行了初步评估。试验结果表明,LLM通过其底层嵌套子模型的协同工作机制,在预测与优化任务中展现出差异化的性能特征。在预测任务方面,LLM在低维简单场景中普遍表现出较高的预测精度,其中DeepSeek模型在低维数据和大样本量条件下展现出显著的性能优势。在优化任务方面,DeepSeek和ChatGPT在大多的测试任务中表现突出。然而,所有测试模型在处理高维优化问题和复杂约束优化问题时均表现出明显的性能局限。旨在揭示LLM在优化设计中的潜力与局限性,为LLM等大模型驱动的工程优化方法提供实践参考。

关键词: 大语言模型, 优化设计, DeepSeek, ChatGPT

Abstract: Large language models (LLM), with their powerful natural language processing capabilities, knowledge integration, and generation abilities, are rapidly transforming research paradigms and practices across multiple domains. However, the potential of LLM in engineering analysis and optimization design has yet to be fully explored. Focusing on the integration of LLM with optimization design. Through conventional experimental design, including nine standard test functions, twelve optimization test functions, and two engineering cases, a preliminary evaluation was conducted on the performance of six mainstream LLM—DeepSeek, ChatGPT, Doubao, Gemini, Grok, and Kimi—in prediction and optimization tasks. The experimental results indicate that LLM exhibit distinct performance characteristics in prediction and optimization tasks through the collaborative working mechanism of their underlying nested submodels. In prediction tasks, LLM generally demonstrate high prediction accuracy in low-dimensional and simple scenarios, with the DeepSeek model exhibiting significant performance advantages under low-dimensional data and large sample conditions. In optimization tasks, DeepSeek and ChatGPT perform exceptionally well in most test tasks. However, all tested models exhibit significant performance limitations when handling high-dimensional optimization problems and complex constrained optimization problems. This study aims to reveal the potential and limitations of LLM in optimization design, providing practical references for engineering optimization methods driven by large models such as LLM.

Key words: large language models, optimization design, DeepSeek, ChatGPT

中图分类号: