Data-Efficient Semiconductor Wafer Defect Identification Using WGAN-GP Augmentation and Lightweight Deep Convolutional Networks
TL;DR
A data-efficient wafer defect classification framework that integrates generative data augmentation with lightweight deep learning models that enables reliable wafer defect classification in data-constrained manufacturing environments, making it well suited for deployment in resource-limited semiconductor fabrication facilities.
Abstract
Automated Visual Inspection (AVI) systems are vital in semiconductor manufacturing for the early detection of wafer defects that degrade yield. However, the deployment of deep learning–based inspection in small and medium-scale fabrication facilities is hindered by limited training data and severe class imbalance, particularly for rare but critical defect patterns. This work presents a data-efficient wafer defect classification framework that integrates generative data augmentation with lightweight deep learning models. A Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) is employed to synthesize realistic defect patterns, with emphasis on underrepresented classes. The quality of the generated samples is validated through visual inspection and distribution-based metrics, including Fréchet Inception Distance (FID) and Kernel Inception Distance (KID). The augmented dataset is used to train and evaluate VGG16, ResNet-50, and a compact YOLOv8-Small classification model. Experimental results on the WM-811K benchmark dataset demonstrate that the YOLOv8-Small architecture achieves superior performance, attaining a Top-1 accuracy of 97.16% and a weighted F1-score of 0.976, while significantly reducing computational complexity. The proposed framework enables reliable wafer defect classification in data-constrained manufacturing environments, making it well suited for deployment in resource-limited semiconductor fabrication facilities.