Machine-learning-assisted analysis of highly transient X-ray imaging sequences of weld pools

Abstract

Fusion-based welding and additive manufacturing are two key pillars of manufacturing. Rapidly evolving melt pools are associated with both of these processing approaches. Understanding and controlling the evolution of the melt pools are critical for optimization of such processes. Flow and interface oscillation during those processes are closely linked to the final fusion zone and microstructure formation. Synchrotron X-ray radiography enables observation of transient melt pools in additive manufacturing and welding processes in real time. However, analysis of the large amount of data generated in such experiments are cumbersome. Thus, we have examined the potential to analyse fast time-resolved X-ray image sequences of melt pools with image-based convolutional neural networks. The results demonstrate successful recognition of changes in the fluctuations of melt-pool interfaces associated with rapid-flow evolution.

Keywords

• Synchrotron X-ray imaging
• weld pools
• fully convolution network
• machine learning