A Selective Convolution Kernel Residual Network for Wafer Map Defect Pattern Recognition

The fabrication of wafers represents a pivotal stage in the production of semiconductors. Defects that emerge during the fabrication process have the potential to result in the production of faulty wafers, which in turn can impact the overall yield of the final product. The analysis of wafer map defect patterns can facilitate the identification of the root cause of defects, thereby enhancing the overall yield of wafers produced. However, the presence of wafer map defects presents a number of challenges, including diversity in location, variability in defect pattern, and size inhomogeneity. Moreover, the current Wafer Map Defect Pattern Recognition (WMDPR) classification algorithm displays shortcomings in terms of accuracy. A Selective Convolution Kernel Residual Network (SCKR-Net) incorporating an Attention Mechanism (Convolutional Block Attention Module, CBAM) is proposed as a solution to the aforementioned problem. The proposed approach employs a Selective Convolution Kernel (SCK) structure, which is capable of adaptively adjusting its parameters and convolution kernel size in response to the varying dimensions of wafer map defects. This adaptability enhances the network's capacity to process complex and variable data. To ascertain the efficacy of the proposed methodology, experiments were conducted on the Mixed-WM38 dataset. The results demonstrate that the proposed method exhibits a markedly superior recognition performance compared to alternative models.