|About this Abstract
||2018 TMS Annual Meeting & Exhibition
||Characterization of Minerals, Metals, and Materials
||Towards High throughput Quantitative Metallography for Complex Microstructures with Deep Semantic Segmentation Models: A Case Study in Ultrahigh Carbon Steel
||Brian L. DeCost, Toby Francis, Elizabeth A. Holm
|On-Site Speaker (Planned)
||Brian L. DeCost
Deep learning methods currently dominate in many visual recognition tasks, but are relatively unexplored and underutilized in microstructure engineering and science.
The complex spatial distribution of grains, phases, and interfaces that we call microstructure can be difficult to define and quantify in a reductive manner, especially for microstructure systems with complex hierarchical structure characteristic of many industrially relevant materials systems. We apply deep convolutional neural network (CNN) architectures designed for image segmentation to label such complex microstructures in SEM micrographs of high carbon steel. We show that CNN models can be trained to produce high-quality maps of complex microstructure features including widmanstatten lath, grain boundary carbide, and cementite particle matrix. These models enable automated statistical characterization of e.g. the width of particle-free denuded zones associated with grain boundaries, paving the way for high-throughput alloy and microstructure design.
||Planned: None Selected