A Prediction Research Of Edge-breaking In Shoulder Process Of CZ Silicon Single Crystal Based On Feature Selection

As the most important semiconductor material,silicon single crystal is widely used in various fields such as national economy and national defense technology.During the growth of CZ silicon single crystals,the shoulder process is the key process to ensure that the crystals enter the equal diameter growth process smoothly.It is easy to have the problem of edge-breaking caused by dislocations,which leads to abnormal growth of crystals.At present,it is mainly for the crystal pulling workers to judge the crystal growth by observing the ridgeline,and there is no effective method.Feature selection is a data preprocessing method.The purpose is to extract features with strong correlation with the target category from the sample set,so as not to reduce or even improve the prediction accuracy of the classifier.It has an important position in the field of data mining and pattern recognition.Therefore,in this paper,a prediction study of the edge-breaking problem in the shoulder process of CZ silicon single crystals based on feature selection is proposed,in order to achieve accurate prediction of the edgebreaking problem and improve crystal pulling efficiency.In this paper,we first introduces the parameters of the single crystal furnace collected in the shoulder process of CZ silicon single crystals and data pre-processing methods.It is proposed to use the correlation coefficient matrix between features to filter the redundant parameters between features,and a sample extraction scheme is proposed.After feature scaling,the training set and test set that are to use in subsequent models are finally obtained.Then in this paper,we use three kind of feature selection algorithms to analyze the correlation between the feature parameters and the edge-breaking in the shoulder process.MIC?MRMR?Relief are used to calculate the correlation coefficients between feature parameters and the edge-breaking problem in the shoulder process,and the calculation results of the algorithms are integrated.The first 30 feature parameters corresponding to each feature selection algorithm are selected as feature subsets to enter the subsequent evaluation study.Finally,in this paper,the classification trainings of the four classifiers are completed in the training set,and we can select the optimal model with the best accuracy and verify the model in the testing set.The 3 features subset are sorted in descending order of correlation coefficients respectively,and the first 6)(6)=1,2,?,30)features are extracted to use as the input features of the classifier for classification training to obtain the highest accuracy corresponding to each feature algorithm on each classifier.After the result comparison,the prediction model with the best accuracy and the smallest feature subset—MRMR-KNN is obtained.The MRMRKNN model is tested on the testing set,and the accuracy is as high as 0.9512,which proves that the model is feasible for the prediction of the edge-breaking in the shoulder process in the growth of CZ silicon single crystal.