Layout Feature Extraction using CNN Classification in Root Cause Analysis of LSI Defects
Layout Feature Extraction using CNN Classification in Root Cause Analysis of LSI Defects
复制标题
LSI 缺陷根本原因分析中使用 CNN 分类的布局特征提取
DOI：
10.1109/tsm.2021.3056717
复制
发表时间：
2021
影响因子：
2.7
通讯作者：
Satoshi Fukumoto
中科院分区：
文献类型：
--
作者：
Yoshikazu Nagamura;Koji Arima;Masayuki Arai;Satoshi Fukumoto
故障的根本原因分析（RCA）是强制性的，以获得LSI的可靠性和生产力。虽然分析布局引起的缺陷是至关重要的优化设计规则和预测未知的缺陷，这是一个具有挑战性的任务，由于难以解释缺陷和电路布局之间的关系。我们应用卷积神经网络（CNN）对LSI布局图像进行分类，以执行先前研究中布局引起的缺陷的RCA。然而，由于图像的低分辨率，实际的缺陷位置不能清楚地区分。在本研究中，我们使用不同大小和分辨率的图像片段进行CNN分类。实验结果表明，提取的布局特征的有效性取决于图像剪辑的分辨率。使用视觉解释技术GradCAM++，当CNN模型在具有更高分辨率的较小图像片段上训练时，可以在局部区域准确地捕获缺陷布局的特征。这些布局特征包括一组图案及其周围环境。相反，利用较小尺寸的剪辑恶化的分类精度，由于从图像的信息较少的合并。在所进行的实验中，即使在使用较小剪辑的情况下，也可以通过增加训练数据集的大小来获得可接受的性能（缺陷位置的检测率DTR> 90%，并且风险图像分类率RCR> 10%）。作为缺陷布局的特征提取的部分布局然后可以用于RCA和设计未来的产品。
Root cause analysis (RCA) of failures is mandatory to obtain the reliability and productivity of LSIs. Although analyzing layout-induced defects is crucial to optimize design rules and to predict unknown defects, it is a challenging task due to the difficulty in explaining the relationship between defects and circuit layouts. We applied convolutional neural networks (CNNs) to classify LSI layout images to perform the RCA of layout-induced defects in a previous study. However, due to the low resolution of images, actual defect positions were not clearly distinguished. In the present study, we use image clips of different sizes and resolutions for the CNN classification. Experimental results indicate that the validity of the extracted layout features depends on the resolution of image clips. Using the visual explanation technique GradCAM++, the features of defective layouts can be accurately captured in local areas when CNN models are trained on smaller image clips with higher resolution. These layout features included a group of patterns with their surroundings. Conversely, utilizing smaller-size clips deteriorates the classification accuracy due to the incorporation of less information from the images. In the conducted experiments, even in the case of using smaller clips, acceptable performance (the detection rate of defect positions DTR ≅ 90%, and the risk-image classification rate RCR ≅ 10%) can be obtained by increasing the size of training datasets. Partial layouts extracted as features of defective layouts can then be used in RCA and in designing future products.