In semiconductor manufacturing, detecting defect patterns is important because they are directly related to the root causes of failures in the wafer process. The rapid advancement of the integrated circuit technology has recently led to more frequent occurrences of mixed-type defect patterns, wherein two or more defect patterns simultaneously occur in a wafer bin map. The detection of these mixed patterns is more difficult than that of single patterns. To detect these mixed patterns, binary relevance approaches based on convolutional neural networks have been proposed. However, as the manufacturing process has been advanced and integrated, various failure types are newly detected, thus the number of single models can be continuously increased following the diversification of defect types. Therefore, we propose an effective framework for detecting mixed-type patterns in which a simple single model, called the single shot detector, is employed. By applying the proposed model to the WM-811K dataset, we show that our framework outperforms existing CNN-based methods and also provides defect location information.

在半导体制造中，检测缺陷图案很重要，因为它们与晶圆工艺中的故障根本原因直接相关。集成电路技术的飞速发展最近导致混合型缺陷图案的出现更加频繁，其中在晶片仓图中同时出现两个或更多个缺陷图案。这些混合模式的检测比单个模式的检测更加困难。为了检测这些混合模式，已经提出了基于卷积神经网络的二进制相关方法。然而，随着制造工艺的进步和集成，新发现了各种故障类型，因此随着缺陷类型的多样化，单个模型的数量可以不断增加。所以，我们提出了一种用于检测混合类型模式的有效框架，其中采用了称为单次检测器的简单单一模型。通过将提出的模型应用于WM-811K数据集，我们证明了我们的框架优于现有的基于CNN的方法，并且还提供了缺陷位置信息。