| Abstract Scope |
Current in-situ monitoring methods for Powder Bed Fusion (PBF) rely heavily on supervised machine learning models trained on specific sensor configurations, material systems, and machine settings. While these models can perform well under controlled conditions, they often struggle when transferred across builds, machines, or defect types. This study investigates the emerging role of Large Language Models (LLMs), particularly vision-language and multimodal models, as reasoning-based tools for interpreting in-situ monitoring data in PBF.
Rather than treating defect detection as a conventional image classification problem alone, this work evaluates whether LLMs can identify and interpret process signatures of defects. The evaluation focuses not only on detection accuracy, but also on the model’s ability to provide physically meaningful explanations and recognize ambiguous or insufficient evidence. The novelty of this work lies in shifting the use of LLMs from general post-process reporting toward process-aware interpretation of real-time signals. |
