Wafer Map Defect Classification Using Autoencoder-Based Data Augmentation and Convolutional Neural Network

TL;DR

A novel method combining a self-encoder-based data augmentation technique with a convolutional neural network to improve the model's generalization capabilities is proposed, offering a reliable solution for wafer defect detection and classification.

Abstract

In semiconductor manufacturing, wafer defect maps (WDMs) play a crucial role in diagnosing issues and enhancing process yields by revealing critical defect patterns. However, accurately categorizing WDM defects presents significant challenges due to noisy data, unbalanced defect classes, and the complexity of failure modes. To address these challenges, this study proposes a novel method combining a self-encoder-based data augmentation technique with a convolutional neural network (CNN). By introducing noise into the latent space, the self-encoder enhances data diversity and mitigates class imbalance, thereby improving the model's generalization capabilities. The augmented dataset is subsequently used to train the CNN, enabling it to deliver precise classification of both common and rare defect patterns. Experimental results on the WM-811K dataset demonstrate that the proposed method achieves a classification accuracy of 98.56%, surpassing Random Forest, SVM, and Logistic Regression by 19%, 21%, and 27%, respectively. These findings highlight the robustness and effectiveness of the proposed approach, offering a reliable solution for wafer defect detection and classification.

Figures (11)

• Figure 1: The architecture of the proposed network model consists of two main modules: the pattern module and the CNN module. The pattern module extracts initial spatial features from the input, which are further processed by the CNN module consisting of two convolutional layers with 5x5 kernels and same padding, followed by max-pooling layers (2x2). Gaussian noise sampled from N(0,1) is introduced to enhance robustness. The processed features are then flattened and passed through fully connected neural network layers with ReLU activation, followed by a final fully connected layer with softmax activation to produce the final class predictions (n3 units). The entire architecture is designed to capture and reconstruct key patterns from the input data.
• Figure 2: Visual Analysis of Wafer Defect Data. (A) Distribution of wafer indices, indicating the frequency of wafers across different index ranges. (B) Sample images of different wafer failure types, including Center, Donut, Edge-Loc, Edge-Ring, Loc (Localized), Random, Scratch, and Near-Full. (C) Distribution of wafer labels, categorizing wafers as "No Label," "Labeled - Non-Pattern," and "Labeled - Pattern" with respective proportions. (D) The frequency of different failure types as a percentage of all pattern-labeled wafers, highlighting the prevalence of various defects.
• Figure 3: Visual Analysis of Wafer Defect Data. (A) Radon transform results of eight different wafer defect types, showing the transformed projections for each type, including Center, Donut, Edge-Loc, Edge-Ring, Loc (Localized), Random, Scratch, and Near-Full. The Radon transform captures the structural characteristics of defects in different directions. (B) Binary mask representations of the same eight defect types, highlighting the most prominent connected defect regions in each wafer map. The yellow regions indicate the detected defect areas after noise filtering, providing a clear view of each defect’s geometric characteristics.
• Figure 4: Comparison of original and reconstructed images using a noised latent vector. (A) The original image. (B) The reconstructed image after adding noise to the latent representation. The visual differences highlight the autoencoder's ability to maintain features despite perturbations.
• Figure 5: Training progress of the autoencoder model based on reconstruction loss. The plot depicts the reduction in reconstruction loss over 30 epochs, indicating improved model performance in reconstructing input images. The gradual decrease in loss reflects the model's increasing capability to capture and encode essential features from the input data.