Developed a deep learning-based framework for automated defect detection and pattern recognition in semiconductor wafer images. Utilized AutoEncoder Convolutional Neural Networks (CNNs) with PyTorch for feature extraction and classification, achieving an accuracy of 98% and a precision of 93% for defect classification. Employed data augmentation techniques and Bayesian Optimization Hyderabad, enhancing model robustness and generalization. Key features of this project are:
-
Developed a deep learning-based framework for automated defect detection and pattern recognition in semiconductor wafer images.
- Designed a pipeline to automate the detection of defects by analyzing high-resolution wafer images. The framework incorporated image preprocessing, feature extraction, and classification to identify and categorize wafer defects, replacing manual inspection methods. The solution was tailored to handle the intricacies of semiconductor data, which often involves detecting subtle and rare defects that can impact yield.
-
Utilized AutoEncoder Convolutional Neural Networks (CNNs) with PyTorch for feature extraction and classification.
- Implemented an AutoEncoder CNN architecture, which was instrumental in learning compressed representations of the wafer images. The encoder captured important defect patterns while the decoder aided in reconstruction, allowing the network to effectively distinguish between defective and non-defective regions. PyTorch was leveraged for its flexibility in building and training these deep learning models.
-
Achieved 98% accuracy and 93% precision in defect classification.
- Through rigorous training and testing, the model achieved high classification performance metrics, with 98% accuracy in identifying defects and 93% precision in categorizing them correctly. This performance demonstrated the model's effectiveness in accurately detecting a wide range of defect types, including subtle, hard-to-detect flaws, thus minimizing false positives and improving overall yield predictions.
-
Employed data augmentation techniques to enhance model robustness.
- Implemented techniques such as rotation, scaling, flipping, and noise injection on the wafer images to artificially increase the size of the training dataset. This not only prevented overfitting but also allowed the model to generalize better to unseen data by simulating various orientations and conditions of defects during the manufacturing process.
-
Applied Bayesian Optimization to improve model generalization and hyperparameter tuning.
- Used Bayesian Optimization to fine-tune the model's hyperparameters, including learning rates, dropout rates, and network architecture parameters. This approach helped identify the optimal configuration for maximizing the model's performance while reducing computation time. As a result, the model demonstrated better generalization across different datasets and manufacturing scenarios.
To get a broad understanding of how this intricate process transforms tiny grains of sand into functional wafers, watch the following short video.
The dataset includes several distinct defect types typically found in semiconductor manufacturing. Here’s a list of the common defect patterns:
- Center: Defects are concentrated at the center of the wafer.
- Donut: A ring-shaped pattern, where defects are found surrounding a relatively defect-free center.
- Edge-Ring: Defects appear around the edge of the wafer in a ring-like shape.
- Edge-Loc: Defects are localized near the wafer’s edge.
- Loc: Localized defects that appear in specific, concentrated regions.
- Random: Defects are scattered randomly across the wafer, with no discernible pattern.
- Scratch: A linear pattern of defects, often caused by physical damage or contamination.
