Deep Learning Approaches for Scanning Electron Microscope Image Analysis of Slurry Coatings
Tuesday, June 10, 2025 3:00 PM to Thursday, June 12, 2025 4:00 PM · 2 days 1 hr. (Europe/Berlin)
Foyer D-G - 2nd floor
Women in HPC Poster
AI Applications powered by HPC TechnologiesEngineeringHigh-Performance Data AnalyticsHPC in the Cloud and HPC ContainersParallel Programming Languages
Information
Poster is on display and will be presented at the poster pitch session.
This study focuses on using deep learning to analyze Scanning Electron Microscope (SEM) images of aluminide diffusion coatings deposited on steel via the slurry route. The goal is to determine the thicknesses of the Fe2Al5 and FeAl coating layers, as well as the characteristics of pores and chromium precipitates. Due to challenges like imaging artefacts, noise, and overlapping features, a deep learning-based SEM image segmentation model using U-Net architecture was developed. Ground truth data were generated with the Weka segmentation plugin in ImageJ, refined manually, and augmented with synthetic data from Blender 3D. The model, trained on both synthetic and real SEM data, achieved mean dice scores of 98.7% ± 0.2 for Fe2Al5, 82.6% ± 8.1 for pores, and 81.48% ± 3.6 for precipitates. The method was applied to evaluate SEM images of coatings from three slurry compositions, revealing that slurries without rheology modifiers resulted in thicker Fe2Al5 layers. Coating thickness did not affect the relationship between outward and inward diffusion Fe2Al5 layers, with thinner coatings showing lower pore and chromium precipitate fractions. A high speed up is achieved by rendering synthetic images on HPC Karolina cluster. Also, a speed of 6.26x is achieved by performing distributed training on 8 GPUs instead of single GPU.
Contributors:
This study focuses on using deep learning to analyze Scanning Electron Microscope (SEM) images of aluminide diffusion coatings deposited on steel via the slurry route. The goal is to determine the thicknesses of the Fe2Al5 and FeAl coating layers, as well as the characteristics of pores and chromium precipitates. Due to challenges like imaging artefacts, noise, and overlapping features, a deep learning-based SEM image segmentation model using U-Net architecture was developed. Ground truth data were generated with the Weka segmentation plugin in ImageJ, refined manually, and augmented with synthetic data from Blender 3D. The model, trained on both synthetic and real SEM data, achieved mean dice scores of 98.7% ± 0.2 for Fe2Al5, 82.6% ± 8.1 for pores, and 81.48% ± 3.6 for precipitates. The method was applied to evaluate SEM images of coatings from three slurry compositions, revealing that slurries without rheology modifiers resulted in thicker Fe2Al5 layers. Coating thickness did not affect the relationship between outward and inward diffusion Fe2Al5 layers, with thinner coatings showing lower pore and chromium precipitate fractions. A high speed up is achieved by rendering synthetic images on HPC Karolina cluster. Also, a speed of 6.26x is achieved by performing distributed training on 8 GPUs instead of single GPU.
Contributors:
Format
On DemandOn Site
