Integrating large models with topology optimization for conceptual design realization

Topology optimization (TO) has been extensively applied in various domains, including robotics, equipment manufacturing, household products, and civil engineering, enhancing the performance of structural designs. However, TO-designed structures often lack adaptability to human preference despite high physical performance. Trained extensively on human knowledge, large visual–language models (LVLM) exhibit a strong ability to understand human intent and generate satisfactory designs efficiently. In this paper, a large visual–language model-guided topology optimization (LMTO) approach is proposed to automatically generate and edit efficient structural designs according to concepts. By integrating the TO into the large model knowledge space through a UDF-Weighting block, LMTO can optimize performance in the direction of human preference. Experimental results show that, despite significant variations in appearance, the performance of the designs remains comparable or superior to those obtained by the BESO method, indicating the effectiveness of our approach in exploring the joint space. Our method can yield diverse designs from the same prompt and is well-adapted to 2D and 3D cases, highlighting its effectiveness and practicality.