Microstructure informatics: Using computer vision for the characterization of dendrite growth phenomena in Ni-base single crystal Superalloys

Abstract

Microstructure informatics, an emerging field, combines traditional quantitative metallography with computer vision, algorithmic geometry and data science. It uses automated procedures to retrieve statistically relevant information from micrographs. Its power is demonstrated in a case study which focusses on competitive dendrite growth during directional solidification of single crystal Ni-base superalloys (SXs) in 3D. We show how microstructure informatics allows to follow the evolution of all dendrites in a cylindric SX bar (diameter: 12 mm, analyzed length: 76 mm), evaluating serial cross sections taken in 1 mm distances. The method presented in this work relies on three basic components: (1) A deep learning object detection network for detecting dendrite core positions. (2) A 3D image reconstruction routine for tracing dendrite paths and (3) a relational geometric ontological (RGO) database, documenting all relevant relationships between individual dendrites. The method allows to characterize crystal mosaicity, individual dendrite growth directions, interactions between dendrites and dendrite deformation. The performance of different deep learning classification architectures (AlexNet, GoogleNet and MobileNetV2) in combination with a YOLOv2 subdetection network is investigated. The network hyper parameters were optimized to achieve detection rates >99 %. A resulting ontological database of 16,631 individual dendrites provides a foundation for further automatic quantitative microstructural characterization.