ReWafer-MultiDefect is an AI system designed for multi-class defect detection in regenerated wafers. This advanced system has been successfully implemented in the production process of a publicly listed company. Utilizing state-of-the-art deep learning technology, ReWafer-MultiDefect accurately classifies various types of defects in wafers, enhancing production efficiency and reducing labor costs.

This project is an improvement upon the previously developed ReWafer-BinaryDefect, which focused on binary defect detection. ReWafer-MultiDefect expands on this by incorporating multi-class classification capabilities to handle a broader range of defect types.

As semiconductor manufacturing demands increasingly precise defect detection, traditional manual inspection methods fall short due to their inefficiency and susceptibility to errors. ReWafer-MultiDefect addresses these challenges by providing a high-efficiency, accurate automated defect detection solution that can classify multiple defect types.

• Challenges:

  • Manual inspections are subjective and error-prone.
  • Traditional methods fail to meet the high precision requirements of modern production.
  • Detecting and classifying multiple types of defects in a single system adds complexity.

• Solution:

  • Develop an AI system for multi-class defect detection with high accuracy.
  • Collect and label data with the help of project personnel to improve model performance.
  • Implement a robust model architecture capable of handling multiple defect categories.

• Results and Future Plans:

  • The system has been successfully deployed, achieving high accuracy in detecting and classifying various defect types.
  • Future plans include enhancing the system to recognize more defect patterns and expanding its application to other areas.

To get started with ReWafer-MultiDefect, follow these steps:

1. Clone the Repository:

   git clone https://github.com/hungcheng-chen/ReWafer-MultiDefect.git
   cd ReWafer-MultiDefect

2. Install Dependencies:

   pip install -r requirements.txt

3. Training: Begin training with the following command. Adjust command-line parameters as needed:

   python train.py --data_dir data/multi_class --num_classes 5 --model_name convnext_tiny.fb_in22k --batch_size 128 --max_epochs 30 --lr 1e-4

4. Validation: Evaluate model performance during or after training with this command:

   python val.py --test --data_dir data/multi_class --num_classes 5 --model_name convnext_tiny.fb_in22k --load_model_path runs/.../best_model.pt

5. Inference: Use this command to detect and classify defects in new wafer images:

   python test.py --test --num_classes 5 --model_name convnext_tiny.fb_in22k --load_model_path runs/.../best_model.pt --image_path .../xxx.png

Due to confidentiality reasons, we cannot provide real wafer images. For this project, we use the convnext_tiny model architecture and employ pre-trained weights from the ImageNet-22k dataset provided by timm for transfer learning.

This project is licensed under the terms of the MIT license. For more details, see the LICENSE file.

• Author: HungCheng Chen
• Email: [email protected]