Defects analysis in LPBF printing based on up-skin and down-skin angles using machine learning

Abstract

In laser powder bed fusion (LPBF) additive manufacturing, geometric features of parts have a significant impact on their printability and quality. Special attention in the literature has been given to the characteristic feature of distinction between up-skin and down-skin surface properties, where up-skin and down-skin surfaces have different orientation angles with respect to the build plate during part fabrication. We focus on slice-level, surface-connected defect analysis relative to up- and down-skin orientation using a YOLO→U-Net pipeline, followed by clustering and statistical morphology. We use a dataset acquired by X-ray CT scanning of LPBF-manufactured Ti–6Al–4V (Ti64) parts and 3D lattice structures with segmented regions of the top and bottom skins. This design allows slice-level analysis of defect geometry with respect to surface-normal direction. Lattice architectures with strut-based and surface-based features with a cell size of 2 mm and strut/wall thickness between 0.25-0.55 mm were utilized in this study. Machine learning and deep learning techniques such as YOLO and U-Net have significantly contributed to the precision and effectiveness of defect and pore detection and classification in 3D-printed components. Outcomes are expected to enhance the defect formation mechanism knowledge and allow optimization of print and design parameters toward improved quality and reliability of key application 3D printed parts.