Accelerated Monitoring of Powder Bed Fusion Additive Manufacturing via High-Throughput Imaging and Low-Latency Machine Learning

Metal 3D printing in particular laser powder bed fusion is in the forefront of product manufacturing with complex geometries. However, these printed products are susceptible to several printing defects mainly due to complexities of utilizing high-power, ultra-fast laser for melting the metal powder. Accurate defect prediction methods to monitor the printing process are of high demand. More critically, such solutions must maintain a very low computational cost to enable feedback control signals for future low-latency laser parameter correction loops, preventing creation of defects in the first place. In this research, first we design an experiment to explore impact of several laser settings on creation of the most common defect called "keyhole porosity". We print an object while recording the laser meltpool with an externally installed high-speed visual camera. After extracting keyhole pore densities, we annotate the meltpool recordings and use it to evaluate performance of a simple but fast CNN model as a low-latency defect detector.