Research On Wafer Map Defect Pattern Recognition Based On Convolutional Neural Network

The process of Integrated Circuit(IC)is very complex;there will be many systematic process problems in the whole manufacturing process.Timely locating and resolving these problems can ensure the yield and IC product quality.Different systematic process problems result in different defect patterns on the wafer map;thus process problems can be located by classifying wafer map defect patterns(WMDPs).This task,also known as wafer map defect pattern recognition(WMPR),has been a hot topic in academia and industry.The WMPR model based on convolutional neural network(CNN)greatly exceeds the WMPR model based on crafted design features in recognition performance.As Moore's Law,More Moore,and More than Moore continue to advance,IC processes are continuously updated;therefore,new systematic process problems and WMPDs are constantly emerging.However,the lack of training samples of new WMPDs makes it difficult to expand the WMPR model which is based on CNN.In view of this problem,an extendable WMPR model based on CNN and few-shot learning is proposed,and this model is experimentally verified on the benchmark data set.The main contributions of this dissertation are as follow:(1)The characteristics of the WMDP are systematically analyzed;the advantages and disadvantages of current methods for WMPR are systematically analyzed,and common crafted design features for WMDP are systematically summarized.(2)Considering the characteristics of the WMDP,a design principle for the CNN used for WMPR is proposed.It is to deepen the convolutional layer while reducing the model parameters.CNN-ENet12,a CNN model used for WMPR and feature extraction network,is designed by following this design principle.On the benchmark dataset,CNN-ENet12 achieves an average recall rate of 94.68 percentage and an average precision rate of 93.84 percentage,which is better than models based on crafted design features and some other CNN models.(3)Aiming at the difficulty of expansion caused by the lack of training samples for new WMDPs,an extendable WMPR model based on combined probability distribution is proposed.The experimental results on the benchmark data set show that the overall performance of this model is reduced by 2.52%?2.88%after each new type of WMDP is extended.Within a certain amount,this model can achieve high-quality extensions.For example,the average F₁ can reach 92.28%after the model being extended one new type of WMDP,which is better than other models.The proposed extendable WMPR model effectively utilizes the ability of CNN to extract the features of WMDPs.Besides achieving high-performance recognition for base WMDPs(which refer to WMDPs with enough training samples),this model can realize high-quality expansion within a certain range.Compared to other models,the extendable WMPR model proposed in this dissertation is more in line with the actual needs of industry application than other models.