January_2021_AMP_Digital

A D V A N C E D M A T E R I A L S & P R O C E S S E S | J A N U A R Y 2 0 2 1 1 9 such as misorientation data and grain or grain boundary visualizations help to better recognize bainitic regions, their boundaries and polygonal ferrite, leading to a more objective and repro- ducible ground truth. Figure 3b shows SEM images of two bainitic microstruc- tures that are to be assigned to the correct class for ML classification. By overlaying the SEM image with EBSD grain boundary visualization and iden- tifying the type of grain boundaries, i.e., sub-grain (red) vs. low-angle (green) vs. high-angle (blue) boundaries, a better is considered. A classification of this carbon-rich second phase was already reported [2,13] . So far, the ground truth assignment has been very manageable as only pearlite, bainite, and martensite had to be distinguished. Now however, bainitic subclasses [14] , e.g., upper and lower bainite are to be added. In theo- ry, upper and lower bainite show quite distinct differences in the type of car- bon precipitation and should be eas- ily distinguishable. However, during the investigation of industrial samples from heavy steel plates, it became clear that there are more degrees of freedom and that in most cases, no typical text- book-like structures of upper and lower bainite were present. In order to obtain clearly defined bainitic structures, spe- cific samples were produced using a quenching dilatometer. See Fig. 4. Isothermal transformation at 525°C yielded fully upper bainite while isothermal transformation at 425°C led to the formation of fully lower bainit- ic microstructures. These benchmark structures can now be used to assign the existing images to upper and low- er bainite. This can be done for exam- ple, by extracting suitable features from the images and performing a similarity search, i.e., finding the images in the ex- isting image set that are most similar to the benchmark structures. This allows the extraction of two image sets for upper and lower bainite, respectively, which then can be used as input for fur- ther supervised classifications. Although in this case study, the process parameters from the sample production directly correlate with the ground truth, the general approach of only using process parameters for as- signing the ground truth must be taken with a grain of salt. It might be tempt- ing to just assign one class for each vari- ation in processing. However, this may cause several problems. First, by “skip- ping” the microstructure analysis, a re- searcher would revert to the approach of empirical processing-properties correlations instead of focusing on a microstructure-based materials devel- opment. Second, different processing routes can yield identical microstruc- tures and identical basic properties. understanding of the microstructure is achieved and an ideal complement to the visual appearance in SEM is found, which helps in performing a more ob- jective ground truth assignment of the bainite class. SPECIFIC PRODUCTION OF REFERENCE SAMPLES In the following example, an ex- isting data set of so-called two-phase steels that consist of a ferritic matrix and a carbon-rich second phase, which is either pearlite, bainite, or martensite, Fig. 3 — Questions regarding region annotation (a) or class assignment (b) can be answered objectively by correlative characterization, i.e., by incorporating EBSD data in addition to the merely visual appearance of the microstructure in LOM/SEM: (c) overlaying LOMwith misorientation parameters and grain visualization from EBSD helps to objectively annotate lath- like bainite regions in complex-phase steel micrographs and (d) overlaying SEMwith different types of grain boundaries from EBSD allows a correct identification of upper and lower bainite.