Machine Learning (ML) assisted image analysis is applied in various fields of material science to gain knowledge about microstructures of different material classes. Since microstructures substantially influence the materials' properties, a deeper understanding of their interaction is necessary to optimize these properties. Even though a manual study of their interplay might be possible for single microstructural parameters, it gets almost impossible if multiple parameters are considered. These parameter restrictions are insignificantly for ML studies, which are also less computationally expensive. In this presentation, I will share two case studies of ML applications for image characterization and analysis. First, I show the analysis of fluorescence images of paper networks. Afterwards, the focus shifts to predicting grain orientations in energy material LiNiO₂ (LNO).

To analyse fluorescence images of paper networks, a pre-trained Mask RCNN (region-based convolutional neural network) algorithm is trained with manually labelled images. Afterwards, the algorithm can predict class, bounding box and mask of multiple instances within the image of interest. The predicted mask is then used to calculate fiber length, width and area, as well as the mean curl ratio and the anisotropy factor of the detected fibers. These are important parameters to generate realistic, digital twins of real paper network microstructures.

A different algorithm is used to predict the grain orientation in an LNO particle. As the first step, diffraction patterns (DP) of an LNO particle are gathered by scanning the particle with transmission electron microscopy. The grain orientations of the DP are determined by template matching and both, DP and corresponding grain orientation, are used to train a convolution neural network (CNN). The algorithm is then capable of predicting the Euler angles (correlate with grain orientation) of a single diffraction pattern. Automizing the analysation of huge amounts of DP, the trained CCN is computationally more efficient, than the extensive template matching previously used.