Week 5 Summary: Unsupervised Learning in Materials

Cross-Book Summary

1. Learning Without Labels

• Why Unsupervised: Most materials data arrives unlabeled; labels are expensive.
• Clustering: K-Means and GMMs discover structure in descriptors and spectra.
• Choosing K: Cluster counts must be defensible, not arbitrary.

2. Embeddings and Autoencoders

• Frozen CNN Embeddings: Pretrained features as unsupervised feature spaces.
• Autoencoders: Bottleneck representations for compression and denoising.
• Anomaly Detection: Reconstruction-error thresholding on nominal data.

3. Application to Microstructures

• Hyperspectral Clustering: Flatten EELS/EDS datacubes, cluster, re-image as phase maps.
• Phase Discovery: Cluster CNN embeddings of micrographs without labels.
• Defect Screening: Flag off-nominal regions via autoencoder reconstruction error.

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.