PSI - Issue 74

Tomáš Vražina et al. / Procedia Structural Integrity 74 (2025) 106 –113 Tomáš Vražina / Structural Integrity Procedia 00 (202 5 ) 000 – 000

107

2

alloying additions, such as carbides or intermetallic precipitates, often compromises ductility and poses a trade-off in mechanical performance (Liu et al., 2018). Laser powder bed fusion (LPBF), an additive manufacturing technique, introduces an alternative approach to strengthening the 316L austenitic stainless steel by promoting the formation of dislocation structures during rapid solidification (Hu et al., 2022). Specifically, the high thermal gradients and cooling rates in LPBF lead to the development of dislocations organized into a shape of hexagonal prism-like structures in 3D (when sliced, the 2D structure is composed of cells) (Kong et al., 2021). These dense dislocation structures act as a barrier to dislocation motion. Therefore higher stresses must be attained for dislocation movement. The resulting strength of the material is increased while maintaining reasonable ductility (Liu et al., 2018; Riabov et al., 2021). Process parameters such as laser power and layer thickness influence the size and morphology of these dislocation cells, allowing microstructural refinement through process control (Kong et al., 2021). However, these finely structured dislocation network is destabilized when an increased temperature is introduced. In-service conditions involving elevated temperatures or cyclic mechanical loads can cause coarsening, annihilation, or rearrangement of the dislocation network, leading to a reduction in yield strength and a drop in fatigue life as investigated by Babinský and Chen (Babinský et al., 2023; Chen et al., 2022). Therefore, rapid and reliable quantification of cell or prism size and density is essential for monitoring microstructural changes during heat treatment and for predicting the performance of additively manufactured parts. (Kong et al., 2021; Wang et al., 2025). Traditionally, techniques based on the line-intercept method have been employed for dislocation cell size assessment. Although considered reliable, these methods are fully manual and extremely time- and labor-intensive. More automated thresholding-based approaches have been explored to reduce manual effort. However, these methods often require extensive tuning of parameters for each image due to differences in brightness, contrast, or image content. Slight variations in image quality may lead to inconsistencies or incorrect results, making such methods unreliable for processing larger datasets without constant user intervention (Stuckner et al., 2022). Recent developments in deep learning have introduced more robust alternatives. Convolutional neural networks (CNNs), particularly U-Net architectures, first proposed by Ronnenberger (Ronneberger et al., 2015), have demonstrated excellent performance in segmenting complex features in medical and materials science imaging (Chaurasia et al., 2023; Mikmeková et al., 2023; Zhou et al., 2024). However, applying them to microstructural features such as dislocation cells in LPBF-processed 316L presents practical challenges. Preparing a training dataset is time consuming, and variations in image contrast can still affect segmentation quality. This study proposes a hybrid image segmentation approach that combines automatic thresholding with deep learning using U-Net. The key idea is to use thresholding-based segmentation as an initial step, which is later manually refined to create accurate training masks. It allows for more efficient preparation of training data without starting annotation from the ground up. The trained model is then used for automated segmentation of new images. This combined approach should reduce annotation time while ensuring high-quality output for complex microstructures.

Nomenclature ANN Artificial Neural Network BIRCH Balanced Iterative Reducing and Clustering using Hierarchies CLAHE Contrast Limited Adaptive Histogram Equalization CNN Convolutional Neural Networks DICE Dice Similarity Coefficient DL Deep Learning EBSD Electron Backscattered Diffraction IoU Intersection over Union LPBF Laser Powder Bed Fusion ReLU Rectified Linear Unit SEM Scanning Electron Microscopy STEM Scanning Transmission Electron Microscopy