Our work investigates the polycrystalline composite deformation behavior through multiscale simulations with experimental data at hand. Since deformation mechanisms on the micro-level link the ones on the macro-level and the nanoscale, it is preferable to perform micromechanical finite element simulations based on real microstructures. The image segmentation is a necessary step for the meshing. Our 2D EBSD images contain at least a few hundred grains. Machine learning (ML) was adopted to automatically identify subregions, i.e., individual grains, to improve local feature extraction efficiency and accuracy. Denoising in preprocessing and postprocessing before and after ML, respectively, is beneficial in high quality feature identification. The ML algorithms used were self-developed with the usage of inherent code packages (Python). The performances of the three supervised ML models-decision tree, random forest, and support vector machine-are compared herein; the latter two achieved accuracies of up to 99.8%. Calculations took about 0.5 h from the original input dataset (EBSD image) to the final output (segmented image) running on a personal computer (CPU: 3.6 GHz). For a realizable manual pixel sortation, the original image was firstly scaled from the initial resolution 10802 pixels down to 3002. After ML, some manual work was necessary due to the remaining noises to achieve the final image status ready for meshing. The ML process, including this manual work time, improved efficiency by a factor of about 24 compared to a purely manual process. Simultaneously, ML minimized the geometrical deviation between the identified and original features, since it used the original resolution. For serial work, the time efficiency would be enhanced multiplicatively.
Keywords: denoising; digital image segmentation; ground truth; machine learning; size spectrum.