Wafer Defect Pattern Classification

In this project, we trained and evaluated various machine learning models to classify defect patterns on semiconductor wafers.

Interactive Wafer Defect Classifier

• Live Demo: https://waferdefects.vercel.app/
• Github https://github.com/Shanthanu-g/waferanalyzer

Dataset

We used the WM-811K Wafer Map dataset for this project

• 811,457 wafer samples from real fabs (172,950 labeled, 638,507 unlabeled)
• Each wafer is a 2-D matrix (0: outside, 1: good, 2: defective)
• 9 classes: None, Edge-Ring, Edge-Loc, Center, Loc, Scratch, Random, Donut, Near-Full

Download

All notebooks have a cell that downloads the dataset and caches it to your computer but if that does not work:

• Download dataset from Kaggle
• Set the file_path= in the notebook to where the dataset is downloaded.

Setup

Step 1 - Install Dependencies

Option A: using uv package manager (recommended)

# Install uv
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install all dependencies and launch Jupyter
uv sync --dev
uv run jupyter notebook

Option B: using pip

python -m venv .venv

# macOS/Linux
source .venv/bin/activate

# Windows
.venv\Scripts\activate

pip install -r requirements.txt

Step 3: Register a kernel

# If using uv
uv run ipython kernel install --user --name=cmpe257-s26-wd

python -m ipykernel install --user --name=cmpe257-s26-wd --display-name "cmpe257-s26-wd"

Step 4 — Run Notebooks

uv run jupyter notebook

Notebooks

1. wafer_defects.ipynb — Exploratory Data Analysis
2. classical_ml.ipynb — Classical ML experiments
3. CNN.ipynb — CNN experiments
4. Autoencoders.ipynb - Autoencoder experiment
5. ResNet.ipynb — ResNet experiment
6. results.ipynb — Compare trained models

Results

Model	Params	Lat. (ms)	Macro F1
Baseline (Random Forest)	—	0.5686	0.652
Wu et al. (XGBoost)	—	0.0338	0.869
CNNSmall	597K	0.4589	0.886
Autoencoder	2.8M	1.0828	0.842
MobileNetV2 ✱	2.2M	—	0.930
CNNLarge	2.4M	1.2167	0.906
CNN + 58f (RF)	2.4M	1.2167	0.906
ResNet-18	11.2M	2.0200	0.862

✱ Literature SOTA; requires separate GAN pipeline