ADVANCED ANOMALY DETECTION IN WAFER MAPS USING DEEP LEARNING

Abstract

Detecting anomalies in wafer maps is, indeed an essential way to ensure quality control in semiconductor manufacturing processes. Deep learning is a field that has become increasingly important for anomaly detection research in recent years. This study evaluates the performance of three deep learning models in advanced anomaly detection: CNNs, Autoencoders, and GANs. Public datasets and industrial repositories provided 200 sample wafer maps, with labeled anomalies of center defects, edge defects, ring defects, and random noise defects. Techniques of preprocessing normalization, noise reduction, and data augmentation were applied to improve model accuracy. Models were trained with cross-entropy loss for classification and mean squared error (MSE) for autoencoders. Optimization techniques were Adam and Stochastic Gradient Descent (SGD). Hyperparameter tuning was done by changing the learning rate, batch size, and model depth. Models were evaluated on accuracy, precision, recall, F1-score, reconstruction error, and ROC-AUC scores. The experimental results indicated that the classifier using GANs achieved the highest accuracy of 95.2% and AUC score of 0.96, while CNN-based models and autoencoders scored 93.5% and 89.7%, respectively, in detecting wafer defects. The autoencoder provided high reconstructions for random noise defects with an MSE of 0.035. The testing of the best model on unseen wafer maps was checked and evaluated by industry experts on practical usability. This study shows that deep learning can significantly improve the detection of wafer anomalies and hence improve the classification of defects, yield management, and process optimization in semiconductor manufacturing. Future work may focus on hybrid models that combine CNNs and autoencoders to further enhance robustness and efficiency.Top of Form