Langmuir. 2025 Jun 16. doi: 10.1021/acs.langmuir.5c01673. Online ahead of print.
ABSTRACT
Plasma Electrolytic Oxidation (PEO) coatings enhance the physical and chemical properties of metallic substrates, including corrosion resistance, wear resistance, and thermal stability. These enhancements are strongly influenced by the porous surface morphology of the coatings, which affects the ion transport, stress distribution, and permeability. Accurate quantification of pore structures is essential for understanding interfacial structure-property relationships, yet traditional image segmentation methods often fail to capture the complexity of PEO surfaces in SEM images. This study presents a deep learning-based segmentation framework using U-Net architectures integrated with Atrous Spatial Pyramid Pooling (ASPP) to improve multiscale feature extraction. The performance impact of ASPP placement within different parts of U-Net was systematically evaluated. Results show that modifications to the bridge and decoder paths have the greatest impact on segmentation performance, with a combined modification applying ASPP in both achieving the highest F1 score (0.9360) and the highest IoU (0.8798). Statistical analysis using 5-fold cross-validation, bootstrap confidence intervals, and paired t-tests confirmed that only the bridge-modified model (B1×1) significantly outperformed the baseline (p < 0.05). The proposed approach enables high-fidelity pore segmentation and supports advanced microstructural analysis of PEO coatings. By facilitating accurate morphological quantification, it contributes to the understanding of structure-property relationships in interfacial materials and offers a robust tool for future materials characterization workflows.
PMID:40523154 | DOI:10.1021/acs.langmuir.5c01673
