Automated Segmentation and Chord Length Distribution of Melt Pools in Complex 3D Printed Metal Artifacts

We present a new computational approach for large-scale segmentation and spatially-resolved analysis of melt pools in complex 3D printed parts and qualification artifacts. Our hybrid segmentation includes human-in-the-loop image processing of a few representative optical images of melt pools that are then used for training machine learning models for automated segmentation of melt pool boundaries in large parts. Our approach specifically targets minimizing the need for manual annotation. Considering imperfect segmentation and errors unavoidable with most algorithms, we further propose chord length distribution as a statistical description of melt pool sizes relatively tolerant to segmentation errors. We first show and validate our new approach on optical images of melt pools in a simple 3D printed plate sample (IN718 alloy) as well as selected regions of a complex qualification artifact (AlSi10Mg alloy). We then demonstrate the application of our approach on an entire cross section of the artifact.