Speaker
Description
Aluminide diffusion coatings are widely applied to steels in high-temperature environments such as steam turbines and solar systems, where they form corrosion-resistant layers that extend component lifetime. Accurate evaluation of these coatings is essential for quality control and accelerated materials development. Manual analysis of Scanning Electron Microscope (SEM) images to quantify features like layer thickness, pores, and chromium precipitates is challenging, time-consuming, and prone to human error due to noise, artifacts, and overlapping features. To address these limitations, this work introduces a deep learning-based approach for automated segmentation and metallographic evaluation of aluminide slurry coatings.
To overcome the limitation of a small manually annotated real SEM dataset, we generated a large corpus of high-fidelity synthetic SEM images together with their ground-truth masks using Blender’s procedural shading tools. Ground-truth labels for real SEM training data were created using the Trainable Weka Segmentation plugin in ImageJ, followed by manual refinement. A U-Net convolutional neural network was subsequently trained on a hybrid dataset combining real SEM micrographs with synthetic SEM images. Training employed high-performance computing (HPC) resources with distributed multi-GPU rendering for synthetic data generation and DistributedDataParallel training to accelerate model convergence. A combined Weighted Dice and Soft Cross-Entropy loss function was found to be most effective in handling class imbalance, particularly for pores and precipitates. Among the tested architectures (U-Net, Attention U-Net, DeepLabV3, and Swin UNETR), the baseline U-Net offered the best performance across most feature classes. The workflow achieved high segmentation accuracy across key microstructural features. On test data, the U-Net model delivered a Dice score of 98.7% ± 0.2 for the Fe2Al5 diffusion layer, 82.6% ± 8.1 for pores, and 81.5% ± 3.6 for chromium precipitates.
In addition to model development, synthetic SEM image generation was benchmarked on a high-performance computing cluster. Parallel rendering reduced synthetic dataset generation time compared to sequential rendering, while multi-GPU training further improved model scalability for larger datasets. Analysis of coatings produced with three slurry formulations revealed that samples prepared without a rheology modifier exhibited thicker Fe2Al5 layers due to dominant inward diffusion. In contrast, thinner coatings contained fewer pores and chromium-rich precipitates, independent of slurry composition.
This study streamlines coating characterization, improves quality control, and accelerates development by providing reproducible, high-accuracy segmentation. By bridging materials science and machine learning, this work demonstrates the potential of synthetic data augmentation and HPC-optimized deep learning workflows for advancing metallographic evaluation of protective coatings.
