AskNatureGPT: an LLM-driven concept generation method based on bio-inspired design knowledge
Liuqing Chen, Zebin Cai, Wengteng Cheang, Qi Long, Lingyun Sun, Peter Childs, Haoyu Zuo
Abstract
Concept generation is the early stage in the engineering design process to produce initial design concepts. By applying bio-inspired design (BID) knowledge, designers can employ biological analogies for solution-driven BID concepts. Solution-driven BID starts with knowledge of a specific biological system for technical design. Despite the proven benefits of solution-driven BID, the gap between biological solutions and engineering problems hinders its effective application, with designers frequently encountering misaligned problem-solution pairs and facing multidisciplinary knowledge gaps in concept generation. Therefore, this research proposes a large language model (LLM) based concept generation method – AskNatureGPT – to automatically search for problems, transfer biological analogy, and generate solution-driven BID concepts in the form of natural language. A concept generator and two evaluators are identified and fine-tuned based on the LLM. The method is evaluated by an ablation study, machine-based quantitative assessments, subjective human evaluations, and a case study. The results show our method can generate solution-driven BID concepts with high quality.