G2LGAN:¹ï¤£¥¿Å¸ê®Æ¶°¶i¦æ¸ê®ÆÂX¼W

À³¥Î©ó´¹¶ê¹Ï¯Ê³´¤ÀÃþ

Abstract:

Semiconductor manufacturing requires multiple complex chemical processes. Errors in any link will cause defects in the produced wafers. Therefore, wafer map defect classification is a key task for the semiconductor industry to maintain and improve yield. However, the current wafer defect classification is still performed by engineers through human labor. Therefore, this paper uses the excellent performance of neural networks in image classification to automatically classify wafer defects. Due to advances in semiconductor manufacturing processes, it is difficult to collect defective wafer maps. In order to solve the problem of data set imbalance and insufficient data, this paper proposes a data augmentation method and random undersampling that can effectively balance the data set. In the part of the wafer map classification network, this article uses a lightweight network architecture, which not only reduces computing resources and improves model efficiency, but also eliminates the problem of overfitting caused by data sets.

The data augmentation proposed in this paper is based on the Generative Adversarial Network (GAN). The proposed method G2LGAN (Global to Local GAN) can first learn the global features of the data set and then learn the local features of individual classes, even in In the case of imbalanced data sets, various classes of data can be effectively generated to solve the problem of insufficient data. The classification network is based on MobileNetV2. According to the experimental results, it is found that in wafer defect classification tasks, low-dimensional features are more important than high-dimensional features. Based on this result, the classification network is designed to reduce the model and maintain high accuracy.

This article uses the WM-811K data set for verification. Because this data set has serious data imbalance problem. This paper integrates data enhancement and random undersampling methods to optimize the data set, and uses the proposed classification network for classification tasks. Experimental results show that the proposed G2LGAN can improve the accuracy of the model by approximately 9.15%. Compared with the existing research, the proposed wafer classification network saves 58% in the amount of model parameters, and has a accuracy of 1.39% and a 12.35% lead in F1-Score.

G2LGAN Flowchart:

Made by ¤ý©Ó·z