Multi-Lable Wafer Defect Pattern Classification Using ResNet-based Convolutional Neural Networks

Wafer testing is a pivotal stage in enhancing semiconductor manufacturing efficiency. Wafer testing is a critical step in determining the quality of chips manufactured on wafers and plays a vital role in chip yield assessment. The results of wafer needle testing are used to generate a Wafer Bin Map(WBM) by analyzing WBMs, enabling the identification of defective wafers. This paper presents an approach for the classification of WBMs. In the preprocessing stage, it employs an adaptive OPTICS-based circular graph filtering algorithm to automatically determine optimal OPTICS parameters. Afterwards, three types of ResNet-based neural network are then employed for Wafer Defect Pattern Classification, resulting in improved accuracy. For the learning rate is 0.01, the classification accuracy reaches the maximum at 97.6%.