Undergraduate Research Assistant

Project Description

This research project investigated the effectiveness of combining computer vision techniques and unsupervised machine learning algorithms for the crystallographic analysis of nanomaterials with grain boundaries. Specifically, the study evaluated the application of Gabor filters for feature extraction from high-resolution Transmission Electron Microscopy (TEM) images and K-means clustering for automated segmentation and classification of grain orientations and defects.

Objective

The primary aim was to develop an image-processing pipeline capable of producing accurate crystal orientation maps and identifying grain boundary defects from TEM images of metallic nanoparticles, without relying on traditional diffraction-based methods or supervised learning models.

Methodology

• Gabor Filters were utilized to detect textural features in TEM images. These frequency- and orientation-sensitive filters were effective in isolating local changes in contrast associated with grain boundaries. A bank of Gabor filters with various spatial frequencies and orientations was created and convolved with the TEM images to extract directional texture information.
• The K-means Clustering algorithm, a centroid-based unsupervised machine learning technique, was applied to the Gabor feature maps. By grouping pixels with similar texture features, the algorithm enabled segmentation of the microstructures into regions corresponding to individual crystal grains or boundaries.

Results

The integrated pipeline demonstrated strong performance in:

• Automated segmentation of polycrystalline structures at the nanoscale.
• Accurate visualization of crystal orientation using false-color mapping.
• Identification of grain boundary defects without the need for manual labeling or extensive training datasets.

Segmentations were successfully applied to metallic systems including ZnO samples, showcasing the potential to detect Σ = 13 grain boundaries and other key features. The approach was validated through visual inspections and comparative analysis across varying cluster parameters (e.g., K=2, K=3, K=6), with colorization enhancing interpretability.

Significance

This work highlights a promising direction for automated crystallographic analysis in materials science. By combining classical signal processing with unsupervised learning, the method avoids the data requirements and biases associated with supervised models while offering high-resolution insights into grain morphology and defect structures. It also opens the door for dynamic, in-situ analysis of grain evolution during phase transformations, especially when paired with aberration-corrected microscopy.

Conclusion

This project successfully demonstrated that combining Gabor filters with K-means clustering provides an effective method for analyzing grain boundaries and crystal orientations in high-resolution TEM images. The approach is data-driven, requires no prior training, and offers a scalable solution for automated materials characterization. These findings highlight the potential for broader applications in nanomaterials research and real-time defect analysis.