ML-based detection of critical defects in additively manufactured parts via X-ray computed tomography

• Significant difference between fatigue-critical and XCT-detected defect sizes. • Voxel size critically limits XCT detection, especially for LoF defects. • U-Net and Triangle segmentation methods improve defect detectability. • ANN-based interaction rule reconstructs fragmented defects accurately. • 3D convex hull approach ensures consistent sizing of critical defects. Volumetric defects in laser powder bed fused (L-PBF) components can critically affect their fatigue performance. Although X-ray computed tomography (XCT) enables volumetric defect detection, its effectiveness is limited not only by physical resolution, but also by post-processing of the data, such as image segmentation. To investigate this limitation, the present study evaluated the influence of different segmentation methods (Otsu, Triangle, and U-Net) on defect detectability and sizing accuracy by analyzing the resulting probability of detection (PoD) and sizing errors across XCT scans with varying voxel sizes. Moreover, to address resolution-induced defect fragmentation, a machine learning (ML) algorithm was used to make informed interaction decisions. In addition, a 3D convex hull method was introduced to reconstruct the size of fragmented defects. Results showed that Otsu often underestimated defect size and missed critical defects at larger voxel sizes, while Triangle and U-Net, especially when combined with ML-based interaction, improved PoD and reduced sizing errors.