Using machine learning to identify in-situ melt pool signatures indicative of flaw formation in a laser powder bed fusion additive manufacturing process

Because many of the most important defects in Laser Powder Bed Fusion (L-PBF) occur at the size and timescales of the melt pool itself, the development of methodologies for monitoring the melt pool is critical. This works examines the possibility of in-situ detection of keyholing porosity and balling instabilities. Specifically, a visible-light high speed camera with a fixed field of view is used to study the morphology of L-PBF melt pools in the Inconel 718 material system. A scale-invariant description of melt pool morphology is constructed using Computer Vision techniques and unsupervised Machine Learning is used to differentiate between observed melt pools. By observing melt pools produced across process space, in-situ signatures are identified which may indicate flaws such as those observed ex-situ. This linkage of ex-situ and in-situ morphology enabled the use of supervised Machine Learning to classify melt pools observed (with the high speed camera) during fusion of non-bulk geometries such as overhangs.