Large Language Models as Few-Shot Defect Detectors for Additive Manufacturing

Additive manufacturing (AM) is a promising tech-nology to tailor complex structures in a layer-by-layer fashion. Its wide adoption, however, is hampered by occurrence of defects that deteriorates product quality. Defect detection methods have therefore been developed to address quality issues, primarily utilizing deep learning models. Despite achieving appreciable results, these models require a substantial amount of labeled data and extensive training, which are costly and time-consuming. Thus, it is worth developing a new approach that can perform defect detection (1) with minimal labeled data, and (2) without the necessity for model updating. Recently, the large language model (LLM) has demonstrated remarkable generalization capability across various fields. It features an in-context learning (ICL) capability that allows it to learn from a context without any retraining. This study explores the potential of applying an LLM for AM defect detection based on ICL. To this end, a dataset pertaining to the resin deposition status in vat photopolymerization, as a field where prevailing LLMs have limited exposure, is collected. Two contributions are made to achieve more accurate defect detection by prompting. First, a multimodal context that includes both image-label pairs and expert knowledge is tailored, which contains comprehensive information about deposition sta-tus. Second, a repeated query and voting (RQV) technique is proposed to mitigate occasional misjudgements, thus enhancing the reliability of the defect detection. Comprehensive ablation studies verify the effectiveness of both the multimodal context and RQV technique. As a result, 96.2% classification accuracy is achieved in such a challenging few-shot setting, obviating the need of model updating.