|About this Abstract
||2021 TMS Annual Meeting & Exhibition
||Additive Manufacturing Fatigue and Fracture V: Processing-Structure-Property Investigations and Application to Qualification
||3-D Convolutional Neural Networks for Pore Analysis in Metal Additive Manufacturing Builds
||Andrew R. Kitahara, Ziheng Wu, Srujana Rao Yarasi, Nihal Sivakumar, Anthony D. Rollett, Elizabeth A. Holm
|On-Site Speaker (Planned)
||Andrew R. Kitahara
Pore structures are of significant interest in metal additive manufacturing (AM) because of their direct relation to mechanical properties. We extend a 2D powder particle characterization tool for application to 3D AM pores. As-built and powder specimens are imaged in 3-D with CT. The pores are segmented from the bulk material, and a pretrained 3-D convolutional neural network (3DCNN) is used as a feature descriptor for the pores. These pores are then clustered via K-Means into a small number of unique morphological types, which are easily verified by human inspection to have similar appearance, shape, size, morphology, and so on. Individual pores can be classified, using machine learning, as intrinsic, keyhole, or lack of fusion porosity. Further, the distribution of pore types can be associated with build parameters. We demonstrate the methodology of our approach and discuss how this analysis tool fits within the framework of exploring process-structure-property relationships.