A Review on the Application of Deep Learning in Wafer Defect Detection

Abstract

Wafer defect detection is a crucial quality control task in semiconductor manufacturing, directly determining the yield and reliability of the chips. With the development of deep learning technology, automated wafer defect detection has evolved from traditional image processing and machine learning methods to advanced neural architectures capable of handling complex defect patterns. This paper comprehensively reviews the deep learning methods used for wafer defect detection, systematically classifying them as classification networks, detection networks, segmentation networks, and emerging hybrid architectures. It also analyzes the evolution from convolutional neural networks to Transformer-based models and visual-language alignment frameworks, highlighting their respective advantages and limitations. Key datasets including WM-811K and MixedWM38, as well as performance evaluation metrics, are introduced. This paper points out key challenges in aspects such as mixed defect identification, class imbalance, model generalization, and real-time deployment constraints. Additionally, future research directions are discussed, covering small sample learning, interpretable artificial intelligence, multimodal fusion, and lightweight model design for edge deployment. This review aims to provide a structured understanding of this field for researchers and practitioners and to guide the development of next-generation intelligent wafer detection systems.