晶圆表面缺陷模式识别的二维主成分分析卷积自编码器

由于半导体制造过程的高度复杂性和动态性,各种过程故障通常导致晶圆表面出现各种缺陷模式.为了有效地识别晶圆表面缺陷模式从而及时地诊断和控制故障源,提出一种深度神经网络模型--二维主成分分析卷积自编码器(two-dimensional principal component analysis-based convolutional autoencoder, PCACAE).首先,提出一种基于改进的二维主成分分析算法(conditional2DPCA,C2DPCA)的图像卷积核,形成PCACAE的第1个卷积层;其次,对卷积输出进行池化操作并卷积编码重构,构建一个卷积编码器,并提取其编码部分作为PCACAE的第2层卷积层的初始化权值,从而形成一个深度网络模型,实现晶圆图像的特征学习;最后, PCACAE网络进行训练微调得到最终网络模型.将PCACAE应用于WM-811K晶圆图像数据库并与其他算法进行对比测试,实验结果表明, PCACAE在晶圆表面缺陷识别上的性能优于其他经典的卷积神经网络模型(如GoogLeNet,DensNet等),从而验证了该方法的有效性与工业可应用性.

Recognition of Wafer Defect Based on Two-Dimensional Principal Component Analysis Based Convolutional Autoencoder

Due to the high complexity and dynamics of the semiconductor manufacturing process, various production faults can result in various wafer defects. In order to recognize the wafer defects and trouble-shoot the root cause of the out-of-control process effectively, a novel deep neural network model, two-dimensional principal component analysis-based convolutional autoencoder(PCACAE), is proposed. Firstly, convolution kernels based on two-dimensional principal component analysis algorithm combined with prior conditions(conditional 2 DPCA, C2 DPCA) are proposed to construct the first convolutional layer of PCACAE. Secondly, the feature maps are pooled and then reconstructed, forming a convolutional encoder. Extract the coding part as weights of the second convolution layer, thus forming a well-pretrained deep network model PCACAE. Finally, fine tune the pretrained PCACAE to get the final model. PCACAE has been successfully applied to the feature learning and pattern recognition of wafer defects. PCACAE is tested in the WM-811 K database and experimental results demonstrated that PCACAE is superior to other well-known convolutional neural network models(such as GoogLeNet, DensNet, etc.) and the effectiveness was proved.