The increasingly demanding requirements profile for steels leads to ever narrower tolerance fields, which must be accompanied by ever more accurate and objective characterization and quantification of the microstructure. This in turn is driving a paradigm shift toward microstructure-based material development. The very fine and complex microstructural constituents, especially in case of quenched and quenched and tempered (Q/QT) steels, hold high demands on suitable sample contrasting and imaging, so that holistic microstructural characterization is coming into focus. At present, however, microstructure evaluation is still dominated by a high degree of subjectivity and strongly depends on the experience of the respective expert. Assessments are therefore often only comparable to a limited extent, which represents a major obstacle for materials development and quality control.

Machine learning (ML) is a promising approach at this point to obtain a reproducible and objective assessment based on quantification that allows direct comparison. A major barrier to the application of ML in materials science and engineering, particularly in microstructural analysis, is a meaningful ground truth assignment. Using correlative microscopy consisting of optical microscopy images, high-resolution microstructural images using scanning electron microscopy, and crystallographic EBSD data that can provide information about local misorientations, this work created a dataset that was then successfully classified using modern deep learning approaches. Due to the complex nature of the investigated Q/QT steels, it is not possible to properly annotate training masks used for conventional semantic segmentation. Thus, a segmentation has been implemented by a classification of representative patches of the different microstructual phases. This allowed the different microstructural constituents to be successfully differentiated from each other in an automated manner with respect to their fine differences in characteristic morphologies. For the evaluation of new unseen microstructural images, a sliding window approach has been used. Hereby, the micrographs have been tiled, using a respective overlap in order to be fed into the classification pipeline. Afterwards, the single tiles have been recomposed and used as representative segmentation result for the respective regions. In this way, a complete segmentation of the entire microstructural image is achieved.