The microstructure of ceramic materials significantly influences their functionality. Traditional microstructural analysis of electron microscopy images of polycrystalline ceramics—complex structures with numerous crystal grains, porosity, and secondary phases—relies on manual analysis by experts, a time-consuming and subjective process. Objective and accurate pixel-based analysis (semantic segmentation) is crucial for quantitative microstructure characterization. While 3D imaging techniques such as optical microscopy, X-ray computed tomography, FIB-SEM, and transmission electron microscopy tomography provide 3D voxel data, analyzing the massive datasets generated requires automated methods. Computer vision approaches, particularly semantic segmentation using machine learning, offer a solution. However, 3D semantic segmentation based on electron microscopy, which offers higher resolution than X-ray techniques, especially for light elements, is less developed than 2D segmentation. This study addresses this gap by applying deep learning to FIB-SEM images of polycrystalline ceramics, focusing on accurate and efficient segmentation despite challenges posed by artifacts and weak contrast in the images.

Existing methods for microstructural image analysis include classical computer vision techniques like thresholding (e.g., Otsu and Sauvola methods) often used with software like ImageJ. These methods have been employed in analyzing diverse functional materials, including superconductors, lithium-ion batteries, thermoelectric materials, and more. However, thresholding methods are limited by their sensitivity to noise and variations in contrast. Advances in computer vision and machine learning, particularly deep learning models such as fully convolutional networks (FCNs) and U-Net, offer significant improvements in segmentation accuracy. FCNs have demonstrated success in image recognition tasks, and subsequent models like U-Net (for medical images) and DeepLab (for autonomous driving) have further refined the approach. While deep learning has shown promise in segmenting microstructures from X-ray computed tomography data, its application to high-resolution electron microscopy images of complex ceramics remains under-explored.

This study employed four semantic segmentation methods: Otsu's thresholding, Sauvola's adaptive thresholding, FCN (FCN-32s, FCN-16s, and FCN-8s), and U-Net. A dataset was created from FIB-SEM images of iron-based high-temperature superconductors. Manual segmentation by experienced graduate students provided ground truth labels. The training images (896x896 pixels) were carefully segmented, with iterative refinement to correct errors and handle ambiguities, particularly near voids where boundary identification is challenging. The depth information from the 3D FIB-SEM data was used to improve accuracy. A separate test image (1100x924 pixels) was used for independent evaluation. Data augmentation techniques (random cropping, rotation, flipping) were employed to expand the training and testing datasets. The deep learning models were trained using BCE Dice Loss as a loss function. The learning rate was adjusted using a decay function, with 120 training epochs. The performance of each model was evaluated using precision, recall, and IoU, derived from the confusion matrix. Finally, 3D reconstruction of the microstructure was performed using the segmented images from the trained models. Quantitative evaluation assessed the variation in the positive phase ratio across z-sections.

The U-Net model achieved the highest IoU (94.6%), outperforming the other methods, including the thresholding methods and FCN variants. The Otsu method yielded the lowest IoU, particularly low recall due to misidentification of noise as defects. Sauvola's method performed better, but showed missegmentations in areas with bright contrast. The FCN models showed improvements in precision with increasing resolution of concatenated features during upsampling. U-Net, concatenating features at all resolutions, demonstrated the highest precision. Qualitative analysis using macroscopic and local views revealed that neural network models were less sensitive to artifacts like ion polishing traces and background intensity variations compared to thresholding methods. While neural networks generally outperformed thresholding, certain challenging structures (impurity phases, shallow voids with bright contrast) resulted in lower accuracy. The 3D reconstruction based on the U-Net and FCN models demonstrated smoother, more continuous structures than those from thresholding, indicating improved accuracy in reconstructing depth information. The quantitative analysis of the filling ratio of the superconducting phase across z-sections showed that the U-Net and FCN-8s models closely matched the manually segmented values, while other methods showed higher variance.

The results demonstrate the superior performance of deep learning models, particularly U-Net, for 3D segmentation of electron microscopy images of polycrystalline materials. The high IoU values achieved are among the highest reported for complex ceramics. The method effectively addresses the challenges associated with artifacts and contrast variations inherent in electron microscopy. The ability to rapidly reconstruct 3D microstructures with high accuracy provides a powerful tool for quantitative analysis of microstructural features impacting material properties, particularly those not easily captured in 2D images (e.g., 3D connectivity of phases, void network structure). This capability significantly advances the understanding of microstructure-property relationships and enables more accurate modeling and simulations.

This study successfully demonstrated the application of deep learning for high-accuracy, rapid 3D reconstruction of electron microscopy images of polycrystalline materials. The U-Net model showed superior performance compared to traditional thresholding and other FCN models. Future work should focus on expanding the training datasets to include a wider range of microstructural features and artifacts to further improve the model's robustness and accuracy, particularly for challenging structures such as impurity phases. The methodology presented here will pave the way for integrating experimental 3D microstructure data with simulations to improve material design and processing.

While the U-Net model achieved high accuracy, certain microstructural features (e.g., small isolated phases, shallow voids) presented challenges. The limited number of impurity phases in the training dataset potentially impacted the model's ability to accurately segment these features. The manual segmentation process, although carefully performed, remains subjective and prone to human error. This is reflected in the fact that the expert human segmented values are only 94.6% accurate.