Machine Learning in microstructure analysis – Why only a holistic, materials science centered approach can form the basis for sustained successful implementations

In the meanwhile, AI and ML have become established in materials science too and have proven their great capabilities and potential. Especially in image processing, i.e., segmentation and classification of microstructures, a wide range of new possibilities is emerging. However, due to the new possibilities, easy availability of ML code and the hype around ML, more and more work is being done that not only lacks a background in materials science, but also leads to questionable approaches and explicit misstatements and misinterpretations.

However, it is no longer sufficient to only consider microstructural images and feed them into an ML algorithm. Instead, the ML-based evaluation must be considered in a holistic approach that also considers all the steps towards obtaining a microstructural image and, with respect to the ML, places a special focus on the assignment of the ground truth. This requires explicit materials science domain expertise. Due to the presented holistic approach - from the definition of the problem, sample selection, metallographic preparation via correlative microscopy to the interpretation of the classification result - ML can either be used to automate comparatively simple questions, where time saving, efficiency increase or relieving the experts from tedious, repetitive tasks is in the foreground, or to enable more complex evaluations and tasks in the first place.

This talk presents a holistic approach to ML-based microstructural analysis, with particular focus on correlative microscopy approaches, and methods for objective ground truth assignment. This is visualized and discussed with various examples from steel research: the segmentation and classification of micrographs, the classification of inclusion types and the classification of fracture surfaces.