Week 5 Summary: Unsupervised Learning in Materials

Cross-Book Summary

1. The Need for Convolutions

• Parameter Explosion: MLPs require too many weights for images.
• Exploiting Structure: Convolutions use shared filters for spatial correlation.
• Shift Invariance: Detects features anywhere in the micrograph.

2. CNN Mechanics

• Convolutional Layers: Hierarchical feature detectors.
• Activation & Pooling: Non-linearity and spatial downsampling.
• Perception Analogy: Inspired by early photo-cell grids.

3. Application to Microstructures

• Segmentation: Pixel-wise classification.
• Object Detection: Bounding specific features.
• Classification: Categorizing whole micrographs.

90-Minute Lecture Strategy

Part 1: Applied Clustering

• Phase segmentation with K-Means/GMMs.
• Defect detection in sensor streams.

Part 2: Applied Autoencoders

• CAE compression of 3D Tomography.
• Latent space defect discovery.

Quarto Website Update (Summary)

Summary for ML-PC Week 5: - Shifts to Unsupervised Learning for unlabeled materials data. - Covers clustering (K-Means, GMMs) and Autoencoders. - Applies techniques to EDS segmentation and sensor anomaly detection. - Explores 3D microstructure latent spaces for automated motif discovery.