Research On Machine Learning And Intelligent Classification&recognition Of Wafer Defect Patterns In Microelectronics Manufacturing

With the development of microelectronics manufacturing technology,semiconductor manufacturing shows the development trend of scale maximization and process-size miniaturization.Even with highly automated processes,high-precision equipment,and advanced technology,it is inevitable that wafers will be abnormal.Abnormal conditions in semiconductor manufacturing process will reduce the yield of wafer products and increase production cost.The later analysis becomes the necessary means to improve the wafer yield.Identification of the wafer can be used to trace abnormal conditions during wafer production.With the rapid development of modern computing power,the trend of applying automatic detection methods based on machine learning to semiconductor production is irresistible.This paper studies the identification and classification methods of 8 common defect pattern types in WM-811 K large-scale industrial data set.The main methods and contents are as follows.Firstly,wafer preprocessing method based on automatic optimized DBSCAN is studied.In addition to the spatial pattern of the wafer map,there are also many noises that can affect the classification of wafer defect types.Therefore,wafer preprocessing is required.Traditional DBSCAN algorithm needs two clustering parameters,and the choice of parameters is easy to affect the clustering effect.Therefore,this paper proposes an automatic parameter filtering method based on DBSCAN,which can solve the traditional drawbacks of manually parameters setting,the algorithm is a Self-Adaptive DBSCAN-based method for wafer bin map,we call it SA-DBSCANWBM.This method selects a comprehensive index of cluster intra-cluster density and inter-cluster density to evaluate the optimal parameters.The experimental results show that the algorithm proposed can automatically and reasonably select better parameters and has a good clustering effect,which is also very helpful for subsequent feature extraction and classification.Compared with the traditional fixed-parameter DBSCAN,SA-DBSCAN wafer map classification accuracy is higher.Secondly,feature extraction and Support Vector Machine(SVM)classification of wafer are studied.In this paper,67 dimensional features are extracted from SA-DBSCAN wafer maps,including 22 dimensional density features,5 dimensional geometric features and 40 dimensional Radon features.Three methods,28 traditional SVM classifiers(One VS One),36 traditional SVM classifiers(28 One VS One and 8One VS Others)and Error Correcting Output Coding SVM(ECOC-SVM),are used to study the defect pattern recognition and classification of these 67-dimensional features in wafer.Among them,the accuracy of ECOC-SVM classifier is 85.3%,higher than that of 28 classifiers and 36 classifiers,and 5.7% higher than that of the WM-811 K dataset provider.Thirdly,Deep Convolutional Neural Network(DCNN)models are constructed.Two DCNN models with different depths are proposed,26-layer Pyramid and 28-layer Pyramid DCNN models.The 26-layer Pyramid DCNN model contains a total of four convolution processing blocks.Three consecutive convolutional layers and pooling layers are used in the first convolution processing block,and two consecutive convolutional layers and pooling layers are used in the second convolution processing block.Using this method can quickly reduce the size of the wafer map feature map and speed up the network training.The 28-layer Pyramid DCNN model contains a total of five convolution processing blocks.In the first two convolution processing blocks,two consecutive convolution layers and pooling layers are used to improve the network training speed while ensuring the accuracy of the classification results.The accuracy of the recognition and classification of the two DCNN models proposed in this article is higher than that of the classic CNN Le Net model and the DCNN models in other articles.Among them,the accuracy of the 26-layer Pyramid DCNN model is 93.2%;the accuracy of the 28-layer Pyramid DCNN model is 93.6%.Fourthly,a combined classification model based on DCNN feature extraction and other classifiers is constructed.The DCNN model can automatically identify and extract the features of the wafer map.Compares the classification results of the DCNN model with the classification results of the ECOC-SVM,KNN and Decision Tree classifiers after the features are derived from the fully connected layer of the DCNN.Using the above three classifiers to classify the features derived from the fully connected layer of the DCNN model has a higher accuracy.Among them,as a multi-classifier,ECOC has error correction capability,can reduce variance and bias,and has high classification accuracy.Therefore,the ECOC-SVM classifier has the best classification accuracy when classifying the features derived from the fully connected layer.The accuracy of using the 26-layer Pyramid DCNN model is 95.7%,and the accuracy of using the 28-layer Pyramid DCNN model is 95.8%.Finally,the classification method combining features of multi-source data with double DCNN model is researched.Uses multi-source data to enrich the wafer map features from different derived images of the wafer map,so that the features include more wafer map features.Drawing lessons from the concept of binocular stereo vision,the two DCNN models mentioned are "binocular".Different DCNN models can obtain different features of wafer maps from different angles.This paper uses the original wafer map and its derived Auto-DBSCAN wafer map and SA-DBSCAN wafer map as multi-source data for feature extraction.For multi-source data corresponding to three groups of features,take the maximum value,minimum value,average value and direct splicing feature classification obtained by these four methods.The experimental results show that the processing effect is better when taking the average value of multi-source data features or direct splicing.On this basis,combining the features of the two proposed DCNN models for secondary processing and classification,the average feature of multi-source data shows greater advantages.The classification accuracy rate is 96.3% when the minimum value of the average feature of the two sets of multi-source data is taken,and the classification accuracy rate when the maximum value,the average value and the direct splicing are taken are all 96.4%.In summary,this article uses multi-source data combined with double DCNN model classification method with an average accuracy of 96.4%.