Classification of bainitic microstructures with machine learning – how to assign the ground truth in the most objective way

The use of machine learning techniques in material science is becoming more and more popular, e.g. in microstructure classification tasks. One of the main advantages cited many times is the better objectivity of these machine learning techniques compared to classification by human experts. However, most of the time the assignment of the ground truth used to build a machine learning classification model is not sufficiently discussed. It can have a subjective component that should not be underestimated, especially when investigating complex microstructures like bainite. Challenges when dealing with bainite are the variety and amount of involved phases, the fineness and complexity of the structures and that there is often no consensus among human experts in labeling and classifying those.

Considering a classification task of distinguishing pearlite, martensite and bainite with its subclasses, this work will show how the accuracy of the machine learning classification model is affected by subjective ground truth assignment when using only SEM pictures. Then methods to assign the ground truth in a more objective way and to improve the classification result will be presented, e.g. the use of reference samples, correlative EBSD data and clustering by unsupervised learning.