|About this Abstract
||2021 TMS Annual Meeting & Exhibition
||Characterization of Materials through High Resolution Imaging
||Using Phase Field Simulations to Train Convolutional Neural Networks for Segmentation of Experimental Materials Imaging Datasets
||Tiberiu Stan, Jiwon Yeom, Seungbum Hong, Peter Voorhees
|On-Site Speaker (Planned)
The ability to quickly analyze large imaging datasets is vital to the widespread adoption of modern materials characterization tools, and thus development of new materials. Image segmentation can be the most subjective and time-consuming step in the data analysis workflow. We show that it is possible to segment experimental x-ray computed tomography (XCT) data of dendritic solidification using a convolutional neural network (CNN) that was trained only using synthetic images. The phase field method is used to rapidly generate these synthetic training images and their associated ground truths without human annotation. The CNN trained on phase field images segmented the experimental data with 99.3% accuracy, comparable to CNNs trained on human-generated ground truths. The number of synthetic images needed for CNN training will be discussed, and the most important microstructural features required for CNNs to “understand” the contents of an image are ranked.
||Machine Learning, Computational Materials Science & Engineering, Characterization