混合模式晶圆图缺陷数据集

晶圆图的缺陷识别对于确定晶圆缺陷成因至关重要，特别是混合模式缺陷。我们在某晶圆制造厂采集了大量的晶圆图数据，这些晶圆图是通过对晶圆片上的每一个晶粒进行电学性能测试得到。但实际采集到各类型的晶圆图在数量分布上存在较大差异，为了保持各类型数据间的平衡，我们采用对抗生成网络生成了部分晶圆图。最终形成了约38000张的混合模式晶圆图缺陷数据集，用于识别混合模式晶圆图缺陷，并辅助晶圆制造工艺中缺陷成因的研究。

Defect recognition of wafer maps plays a critical role in determining the root causes of wafer defects, especially for mixed-mode defects. We collected a large volume of wafer map data from a wafer fabrication plant. These wafer maps were generated through electrical performance tests performed on every die on the tested wafers. However, the collected data exhibited substantial imbalance in the sample counts across different defect categories. To balance the dataset across all types, we utilized Generative Adversarial Networks (GANs) to synthesize additional wafer map samples. Ultimately, we developed a mixed-mode wafer map defect dataset containing approximately 38,000 samples. This dataset is designed for mixed-mode wafer map defect recognition tasks and to assist in research on the root causes of defects in wafer manufacturing processes.

搜集汇总

背景与挑战

背景概述

该数据集是一个用于混合模式晶圆图缺陷识别的专业数据集，包含约38000张晶圆图，旨在辅助晶圆制造工艺中缺陷成因的研究。数据来源于实际制造厂采集和对抗生成网络合成，以平衡不同类型缺陷的样本分布，每张图以52×52数组形式表示，编码了空白点、正常晶粒和失效晶粒，并配有8种基本缺陷类型的one-hot标签。