In semiconductor manufacturing, pattern analysis of wafer maps is important in terms of failure analysis and activities to increase yield. Image classification research using deep learning becomes popular; its application in wafer map classification in semiconductor manufacturing is also growing. However, to improve defect analysis accuracy, through-wafer map classification and clustering, more accurate pattern classification and data processing methods are required. It is difficult to identify the wafer map data expressed in the form of hundreds or thousands in dozens of patterns, and the frequency of the wafer map shape varies according to the change in yield. We present a learning method of a wafer map classifier that can process data of an undefined pattern without compromising the classifier's accuracy and evaluate its performance by applying the learning method to a model widely known in image classification. The data used in our study uses real wafer map data.

在半导体制造中，晶圆图的图案分析对于失效分析和提高产量的活动很重要。使用深度学习的图像分类研究变得很流行；它在半导体制造中晶圆图分类中的应用也在增长。但是，为了提高缺陷分析的准确性，晶圆图的分类和聚类，需要更准确的图案分类和数据处理方法。难以识别以几十个图案中的数百个或数千个形式表示的晶片图数据，并且晶片图形状的频率根据成品率的变化而变化。我们提出了一种晶圆图分类器的学习方法，该方法可以处理未定义图案的数据而不会损害分类器的性能。通过将学习方法应用于图像分类中广为人知的模型来提高其准确性并评估其性能。我们的研究中使用的数据使用真实的晶圆图数据。