Facilitating design for additive manufacturing with KG-based retrieval-augmented generation

Wan, Yuwei 2025. Facilitating design for additive manufacturing with KG-based retrieval-augmented generation. Presented at: 31st International Conference on Engineering, Technology, and Innovation- ICE IEEE/ITMC, Valencia, Spain, 16-19 June 2025. Item availability restricted.

PDF - Accepted Post-Print Version Restricted to Repository staff only until 9 July 2025 due to copyright restrictions. Download (697kB)

Abstract

Additive manufacturing (AM), or 3D printing, enables the production of complex geometries and highly customized components. However, design for AM (DfAM) requires specialised and comprehensive knowledge of process constraints, material behaviours, and performance parameters. While knowledge graphs (KGs) have been utilized to organize and integrate DfAM knowledge, they do not understand natural language and have limited reasoning capabilities, which make them less accessible to non-experts and ineffective in handling complex and context-dependent design queries. Large language models (LLMs) offer powerful language processing and generalizability but suffer from hallucinations when handling specialized domains. To address these limitations, this study proposes a KG-based retrieval-augmented generation (RAG) approach to develop a domain-specific question-answering (Q&A) in DfAM. By integrating structured knowledge from a DfAM KG with the strengths of LLMs, the proposed approach improves response accuracy and relevance. Comparative experiments evaluated LLMs with non-RAG and KG-based RAG using generic and domain-specific metrics. Results demonstrated that KG-based RAG enhances information retrieval and response quality, reduces hallucinations, and ensures alignment with domain knowledge.

Item Type:	Conference or Workshop Item (Paper)
Status:	In Press
Schools:	Schools > Engineering
Date of First Compliant Deposit:	30 May 2025
Last Modified:	09 Jun 2025 14:15
URI:	https://orca.cardiff.ac.uk/id/eprint/178608

Actions (repository staff only)

Edit Item