Grain size determination is crucial for characterizing materials, particularly ceramics, as their

microstructure impacts properties and performance of the material. Manual methods like line-intercept

counting are time-consuming and error-prone, with deviations up to ±10% depending on the number

of grains evaluated, the number of phases and the homogeneity of the material. [1]. This study applies

a U-Net Algorithm to SEM images of Al2O3 & ZrO3 ceramic material, demonstrating its effectiveness.

In this study, a U-Net model was trained to segment grain boundaries in SEM images of Al₂O₃ and ZrO₂

ceramics. The dataset comprised 30 images (2048×1536 px); 15 were used for training (90%/10% split),

and 15 for final validation. Grain boundaries were manually labeled and binarized. All images were

divided into 512×512 patches, and data augmentation (rotation, mirroring) increased the dataset size

eightfold. The network architecture consisted of five layers with 1.9 million parameters.

Performance was evaluated using IoU and Dice scores, and compared against manually segmented

validation data per ISO standards [1]. Average deviation in grain size was 4.5%. Classical algorithms (e.g.,

watershed) were applied to the binary images for grain size extraction. Histogram-based grey-value

thresholding allowed phase-specific grain size analysis (Figure 1, right).

The neural network provides a consistent assessment, reduces subjective bias and offers a reliable

alternative to manual segmentation. In addition, more data can be obtained in less time. New

characteristics of the grains, such as aspect ratio or orientation, could also be analysed without

additional effort. Histograms, as shown in Figure 1 on the right, are also very difficult to produce

manually. Visual evaluation remains important, as IoU and Dice scores do not always align with visual

impressions.