Deep learning accelerated micrograph-based porosity defect quantification in additively manufactured steels for uncovering a generic process-defect-properties relation

Abstract

To date, the micrograph-based defect examination remains one of the primary methods for evaluating the quality of additively manufactured (AM) components due to its low cost and unique advantage in providing direct and precise observations of existing defects. However, traditional micrograph-based defect examination typically relies on human operation, which largely restricts its efficiency and repeatability owing to subjective manual intervention, thereby impeding the accurate and rapid evaluation of defects in large-scale specimens. In this study, deep learning (DL) is employed to accelerate the micrograph-based defect examination via training a semantic segmentation model for defect recognition and quantification, contributing to enhance both the efficiency and precision of this method by replacing conventional manual operations. The proposed DL method is successfully applied to 50 laser powder bed fusion (L-PBF) specimens of 18Ni300 maraging steel to rapidly and accurately recognize two kinds of porosity defects, i.e., Gas-Entrapped Pore (GEP) and Lack of Fusion (LoF), in 5000 micrographs and then provide reliable quantification outcomes. Furthermore, a generic relation among printed parameters, porosity information, and tensile properties is meticulously investigated based on these large-scale quantitative defect results. Besides, this work also provides a detailed discussion of the trained model's robustness to image quality and sample quality, as well as the impact of the observation area on the quantification results.