Capturing Experiential Sustainability Design Knowledge through an LLM-Based Acquisition Module for KBE Applications

This paper presents the design and evaluation of an LLM-based knowledge-acquisition module that converts free-form, experiential design heuristics for sustainable product development into machine-readable records for knowledge-based engineering (KBE). Heuristics are formalized with a domain-specific meta-model and grounded to the eClass vocabulary via retrieval augmented generation (RAG). The prototype combines in-context prompting of a selected LLM (GPT-4o-mini) with a RAG layer (bge-large-en-v1.5). On an unseen student corpus of 44 heuristics (473 labels), the system achieved precision of 0.96, recall of 0.90, and an F1-score of 0.93. A qualitative error analysis attributes remaining recall losses mainly to conservative effect extraction. The results indicate that LLM-driven extraction can address the knowledge-acquisition bottleneck in KBE at compatible cost and latency, while highlighting threats to validity, sensitivity to model drift, and pathways toward ontology integration and simulation-based validation. Prompts, source code, and the aggregated test data set to support replication are published with the paper.