Automated defect characterization and physics-based performance modeling of additively manufactured metals with vision transformers

Abstract Machine Learning (ML) algorithms can optimize data classification, facilitating automated defect detection through microscopy and predicting fatigue performance based on defect levels. This report describes our study in which we implemented Vision Transformers (ViTs) to characterize defects and investigate defect-induced fatigue behavior of Direct Metal Laser Sintering (DMLS)-printed 316L stainless steels. We performed the Moore fatigue test to evaluate the dynamic behavior of the steel samples. Scanning electron (SEM) and optical microscopy (OM) images of 316L steels were collected to map the cross-sections of the samples. We trained our machine vision model to recognize and classify the main types of AM defects based on their location, size, morphology, and type. The results quantified the relation between the defect topology, location, and fatigue performance of the 316L samples. ML-assisted automated characterization of DMLS steels offered a high-throughput approach to establish the relationship between microstructural attributes and macro-scale behavior in AM metals. Graphical abstract