Learning Moisture-Induced Damage From Vision: Diffusion Models for Real-Time Monitoring of Additive Manufacturing Processes.

Moisture is a subtle but critical factor in polymer manufacturing, particularly in additive manufacturing (AM) processes. Many polymers, including thermoplastic polyurethane, are highly hygroscopic and readily absorb ambient moisture, which can lead to defects such as stringing, pores, and bubbles. These defects degrade both print quality and mechanical performance, posing a significant challenge for AM as a reliable next-generation manufacturing technology. Due to these challenges, real-time monitoring and integrity estimation of fabricated parts have become essential to ensure manufacturing quality control. Here, we create an in situ visual monitoring system for fused filament fabrication using an optical camera-based setup to detect moisture-induced degradation and evaluate the quality of printed parts. A diffusion model-based anomaly detection framework is devised to identify the degradation. Our model can identify filaments affected by moisture and assess the extent of degradation from captured images. Furthermore, the system demonstrates that the detected anomaly score is closely correlated with the mechanical performance of the printed parts, offering a nondestructive evaluation approach. These results show that an integrated visual monitoring system and generative artificial intelligence models can provide a robust foundation for enhancing the reliability of additive manufacturing and support resource-efficient sustainability through early, nondestructive detection of defects.