Disclaimer: Certain details (e.g., exact data volume, ratio of normal vs. defect samples) are approximated for industrial confidentiality.

Purpose

Automatically identify wafer defects early in the 3D NAND manufacturing process, reducing manual inspections and improving yield.

1. Introduction & Background

Target: Identify minor scratches, pattern irregularities, or foreign particles on wafers.

Motivation: Traditional manual inspection is time-consuming and error-prone, particularly for high-density 3D NAND structures.

Approach: Use a Convolutional Neural Network (CNN) to classify wafer images as "Normal" or "Defect Suspected."

2. Data & Labeling

2.1 Data Sources

• Inspection Cameras: High-resolution (4K) images captured at multiple stages of wafer production.
• Historical Inspection Records: Human operators' logs indicating defect locations and types, used for ground-truth labeling.

2.2 Data Volume

~25,000 wafer images gathered over 3 months, with a roughly 10:1 ratio of normal-to-defective wafers.

We balanced the dataset partly through strategic sampling and data augmentation.

2.3 Label Annotation

• Each image is labeled as "Defect" or "Normal" based on final QA outcomes.
• For advanced analysis, some images include bounding boxes highlighting the defect region.

3. Preprocessing & Data Augmentation

3.1 Preprocessing

• Resolution: Resized all wafer images to 224×224 (maintaining aspect ratio).