Wafer Maps Want to Be Images

A wafer map (the grid of pass/fail results for every die on a wafer) is usually treated as tabular data. This note argues that's a category error. A wafer map is an image, and the moment you treat it as one, fifty years of computer vision becomes applicable.

The experiment

Take die-level bin maps, render them as small single-channel images (die state → pixel intensity), and ask a standard vision pipeline to do three things:

1. Classify known spatial signatures: edge rings, center clusters, scratches, radial patterns.
2. Cluster unlabeled maps to discover recurring signatures nobody has named yet.
3. Retrieve historical wafers with similar patterns, because a process engineer's first question is always "have we seen this before?"

What worked

• Even a small CNN embeds wafer maps into a space where cosine similarity matches engineer intuition about "same problem."
• Rotation augmentation matters more than architecture: many signatures are rotationally equivalent because the physical cause (e.g., a chamber asymmetry) can present at any angle.
• The retrieval use case beat classification for engineer adoption. Engineers didn't want a label; they wanted precedent.

What didn't

• Pretrained ImageNet backbones added little. Wafer maps are too far from natural images; training from scratch on a few thousand maps did better.
• Autoencoder anomaly scores flagged rare patterns, not important ones. Rarity and severity are different axes, and conflating them produces alert fatigue.

Next

Pairing map embeddings with process step metadata, so the system can not only say "this looks like an edge ring" but "edge rings like this historically correlated with step 47."