Segmentation and Metallographic Evaluation of Aluminium Slurry Coatings Using Machine Learning Techniques

Analysis of scanning electron microscope (SEM) images is crucial for characterising aluminide diffusion coatings deposited via the slurry route on steels, yet challenging due to various factors like imaging artefacts, noise, and overlapping features such as resin, precipitates, cracks, and pores. This study focuses on determining the thicknesses of the coating layers Fe₂Al₅ and, if present, FeAl, pore characteristics, and chromium precipitate fractions after the heat treatment that forms the diffusion coating. A deep learning SEM image segmentation model utilising U-Net architecture is proposed. Ground truth data were generated using the trainable Weka segmentation plugin in ImageJ, manually refined for accuracy, and supplemented with synthetic data from Blender 3D software for data augmentation of a limited number of SEM label images. The deep learning model trained on a combination of synthetic and real SEM data achieved mean dice scores of 98.7% ± 0.2 for the Fe₂Al₅ layer, 82.6% ± 8.1 for pores, and 81.48% ± 3.6 for precipitates when evaluated on manually labelled SEM data. The deep learning procedure was applied to evaluate a series of SEM images of diffusion coatings obtained with three different slurry compositions. The evaluation revealed that using a slurry without a rheology modifier may lead to a thicker partial Fe₂Al₅ layer that is formed by inward diffusion. The relation between the outward and inward diffusion Fe₂Al₅ layers was not affected by the coating thickness. The thinner diffusion coating presents lower pores and chromium precipitate fractions independently of the slurry selected.