In recent years, machine learning has been established as an important tool for microscopists and material scientists. Widespread use of these data-driven approaches can be observed in academic work as well as industry. In this work, we want to highlight several case studies of machine learning usage in a typical material characterization workflow in microscopy. First, we demonstrate the importance of image quality for materials analysis and how deep learning can help to evaluate and improve it. Focus and motion blur can be major challenges for reliable quantitative microstructure analysis, especially in automated microscopy workflows. Examples of reconstructing microstructure information from blurred data will be shown on examples of 100Cr6 with lower bainite and martensite, AlSi casting alloys, sintered FeNdB magnets, and Li-ion batteries, among others. [1,2] The second case study shows how machine learning can aid and automate quantitative microstructure analysis for automated grain size analysis.  Deep learning models can be applied to segment grain boundaries for a wide variety of materials like copper, austenite, brass, aluminum, and others, while handling even complex cases where the grains show substructures that would interfere with classical image analysis methods. Similar models are further applied to implement a material-independent postprocessing to correct incomplete grain boundaries. [3] Defects and local microstructural deviations can have a big impact on materials properties, product safety, and aging behaviour. We showcase different supervised and unsupervised data-driven approaches to detect such defects in Li-ion batteries and FeNdB magnets. We also show how the characteristics of the defect population are decisive in selecting the right methods and how to bridge the gap between detection and quantification. [4] For the last case study, we demonstrate machine learning for predicting materials properties directly from micrographs. This is shown for Li-ion battery electrodes with different porosities. The use of explainable artificial intelligence methods will be introduced to gain insight into the model predictions and how they correspond to the actual microstructure. [5]

References
[1] P. Krawczyk International Conference on Automation Science and Engineering, 2021, pp. 1332-1337.
[2] P. Krawczyk Practical Metallography, 2021, vol. 58, no. 11, pp. 684-696.
[3] K. Rathod Materials Characterization, 2024, 217, 114379.
[4] A. Jansche FEMS EUROMAT, 2023.
[5] P. Deeg Batteries, 2024, 10, 99.