SiMiC: Context-Aware Silicon Microstructure Characterization Using Attention-Based Convolutional Neural Networks for Field-Emission Tip Analysis

TL;DR

The paper tackles the bottleneck of SEM-based silicon microstructure measurement by introducing SiMiC, an attention-enabled CNN that automatically extracts geometry (width, height, apex radius) from SEM images and relates it to field-emission performance. The framework combines a CNN backbone with a CoordConv-based structure module and attention mechanisms to fuse geometric cues, evaluated on a 900-image silicon tip dataset with robust data augmentation and multiple backbones. Key findings show ResNet as the strongest baseline, multi-head attention delivering the best accuracy, and that augmenting data plus providing explicit geometry inputs markedly improves radius prediction, though overall correlation remains modest. The work provides a public SIMIC dataset and baseline code, offering a data-driven path toward predictive emitter engineering and faster, reproducible microstructure analysis.

Abstract

Accurate characterization of silicon microstructures is essential for advancing microscale fabrication, quality control, and device performance. Traditional analysis using Scanning Electron Microscopy (SEM) often requires labor-intensive, manual evaluation of feature geometry, limiting throughput and reproducibility. In this study, we propose SiMiC: Context-Aware Silicon Microstructure Characterization Using Attention-Based Convolutional Neural Networks for Field-Emission Tip Analysis. By leveraging deep learning, our approach efficiently extracts morphological features-such as size, shape, and apex curvature-from SEM images, significantly reducing human intervention while improving measurement consistency. A specialized dataset of silicon-based field-emitter tips was developed, and a customized CNN architecture incorporating attention mechanisms was trained for multi-class microstructure classification and dimensional prediction. Comparative analysis with classical image processing techniques demonstrates that SiMiC achieves high accuracy while maintaining interpretability. The proposed framework establishes a foundation for data-driven microstructure analysis directly linked to field-emission performance, opening avenues for correlating emitter geometry with emission behavior and guiding the design of optimized cold-cathode and SEM electron sources. The related dataset and algorithm repository that could serve as a baseline in this area can be found at https://research.jingjietan.com/?q=SIMIC

Figures (3)

• Figure 1: Overview of the model architecture. The CNN backbone (blue rectangle) extracts features from the input image to predict target parameters. The optional structure module (green rectangle) provides reference dimensions to align the output, while the attention module (red rectangle) facilitates interaction between the backbone and structure features.
• Figure 2: The figure presents the original images alongside their augmented versions, generated through systematic variations in brightness ($\beta = -40, 10, 60$) and contrast ($\alpha = 0.6, 1.1, 1.6$), including the width (W), height (H), and radius (R). These augmentations aim to simulate varying lighting conditions due to different exposure levels and adjustments by SEM.
• Figure 3: Attention map visualizations comparing two ResNet-based prediction settings: (a) Half Prediction, where object height and width are provided as input features, and (b) Full Prediction, where these geometric cues are absent. Red regions indicate high attention weights, while blue areas reflect minimal focus. When height and width are omitted, the model allocates greater attention to the object's bounding structure, implicitly inferring geometric attributes. Conversely, providing these features leads to more localized attention on object contours, improving radius estimation. Notably, background areas also receive attention, suggesting their utility as contextual anchors for object scale and positioning. The sample images were randomly picked from the dataset for visualization.