Porosity segmentation in X-ray computed tomography scans of metal additively manufactured specimens with machine learning

• Published in: Additive manufacturing Vol. 36; p. 101460
• Main Authors: Gobert, Christian, Kudzal, Andelle, Sietins, Jennifer, Mock, Clara, Sun, Jessica, McWilliams, Brandon
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2020
• Subjects: Convolutional neural networks; OTSU thresholding; Powder bed fusion; X-ray computed tomography; Powder bed fusion; X-ray computed tomography; Convolutional neural networks; OTSU thresholding
• ISSN: 2214-8604; 2214-7810
• DOI: 10.1016/j.addma.2020.101460

X-ray computed tomography (XCT) is widely used in additive manufacturing (AM) to obtain discrete analysis of internal material discontinuities, especially the porosity of AM specimens. XCT uses X-ray penetration to generate 3D digital reconstructions that enable non-destructive evaluations of specimens and their internal structures. The process of segmenting XCT images for porosity analysis can be time consuming, affected by XCT scan quality, and subject to segmentation methods. OTSU thresholding and a convolutional neural network were combined into a machine learning tool to automatically segment porosity from XCT images of metallic AM specimens. Multiple XCT specialists and AM specimens were used to investigate how various segmentation methodologies, used to create ground-truth labels of porosity, impacted machine learning performance. XCT specialists segmenting a control specimen established a benchmark for machine learning performance measured through classification and descriptive statistics. Discrepancies in the machine learning tool segmentations were similar to or better than the discrepancies among the XCT specialist themselves, indicating a high capability for automated porosity segmentation.