A Novel Deep Learning Architecture for Global Defect Classification: Self-Proliferating Neural Network (SPNet)

The competition in the semiconductor industry is intense, and any manufacturing company's primary concerns are to reduce costs and improve quality and reliability. By increasing the yield and maximizing the throughput of good wafers, lower costs and higher revenues can be achieved. The ability of defect inspection affects the product yield and productivity. The high rate of false negatives in defect inspection will result in defective wafers being treated as normal wafers and shipped to customers. In addition, a high false positives rate will cause non-defective wafers to be considered abnormal and lead to additional review loading by engineers. Therefore, how to reduce both false negatives and false positives is the main challenge for defect inspection. In this paper, we have developed a new deep learning architecture, named Self-Proliferating Neural Network (SPNet). Compared with other methods, SPNet can significantly reduce false positives and false positives, while improving quality and productivity. We also show that our method generalizes well to other public datasets, where they achieve state-of-the-art results. Finally, we apply SPNet to the classification tasks of defect map and defect pattern, and the F1-score achieves 98.9% and 98.2%, respectively. We conduct experiments that probe the robustness of learned representations and conclude that SPNet has significant benefits in robustness and generalization.