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 2 0 Finally, it can lead to a definition of too many classes and thereby introduce the danger of overfitting the ML model. UNSUPERVISED LEARNING TECHNIQUES In contrast to supervised learning, there is no assignment of the ground truth by a human expert during unsu- pervised learning, whichmeans that the ML algorithm is allowed to find struc- tures in its input on its own. The most widespread unsupervised learning technique is clustering, which is used to explore data and find patterns or to dis- cover groups of similar data points. A general pipeline for cluster- ing images by unsupervised learn- ing was suggested by Kitahara et al. [15] . It consists of feature extraction by a pre-trained convolutional neural net- work, followed by dimensional reduc- tion (principal component analysis plus t-distributed stochastic neighbor em- bedding (t-sne)) and finally a clustering algorithm (k-means). Of course, other types of features can also be used as in- put for this pipeline, e.g., manually en- gineered features from conventional microstructure quantification. For this case study example, two- phase steels consisting of a ferritic ma- trix and a carbon-rich second phase that is either pearlite, bainite, or marten- site are considered again and EBSD pa- rameters (standard pattern quality and misorientation parameters), extracted separately for every second phase ob- ject, are used as input features. The first step investigates whether clustering can find the three main classes (pearl- ite, bainite, or martensite) in the data. Figure 5a shows the resulting clusters (after dimensional reduction (t-sne), determination of the optimal cluster number, and fuzzy c-means clustering). Black circles indicate data points with a low cluster membership grade. These are “low confidence data points” in the fuzzy overlap between clusters 1 and 2 or too far away from the cluster centers and can be filtered out before further evaluation. The corresponding human expert labels are shown in Fig. 5b. They are almost completely consistent with the clustering result (97.3%agreement). In current research, this clustering approach is extended to bainite sub- classes, enabling the comparison be- tween human expert labels plus class definitions and the unbiased, AI de- termined clusters and introduce more objectivity to the controversy of bain- ite classification. Moreover, unsuper- vised learning provides another general benefit: the better insight into and un- derstanding of the input data as it is possible to investigate why certain data points are close to the cluster centers, Fig. 4 — Specific sample production using quenching dilatometer (1) yielding benchmark structures for upper (2) and lower bainite (3). These can be used to divide real bainitic objects from industrial heavy plates (4) into upper (5) and lower bainite (6) partition. Fig. 5 — (a) 2D t-sne plot with 3 resulting clusters after fuzzy c-means clustering. (b) 2D t-sne plot with human expert labels.