Machine learning-Enabled Multi-scale Process Monitoring and Development for Metal Additive Manufacturing

 The variability in the printing outcome of additive manufacturing of metals is a major obstacle that hinders the reliance on the quality of printed parts and thus the potential for full production. In-situ monitoring coupled with machine learning can save labor-intensive costs and time-consuming ex-situ work by enabling accelerated process design that targets consistent printing quality. To that end, this dissertation proposes smart monitoring approaches at the melt-pool scale and part scale through high-speed imaging and IR thermal imaging for the process development, identification of defects formation, and analysis of the humping-up phenomenon as one of the major surface defects.

We start by looking into the global effect of heat accumulation and monitoring the process at the part scale by using IR thermal imaging as the most convenient thermal monitoring. An unsupervised machine learning-based method is developed to detect heat accumulation in real-time without the need to spend a long time labeling millions of images. Furthermore, we have developed an approach to generate thermal distribution given the laser toolpath by using generative deep learning. We have investigated the effect of the parameters of the scanning strategy on thermal distribution. Finally, we zoom into the process and track the molten pool of metal as the laser moves to dive into the local aspects of the process.