Research On "Broken Edge" Prediction Method For Equal-diameter Growth Process Of Single Crystal Silicon Based On Data Mining

The process of Czochralski single crystal silicon growth is a very complicated.Various sensors need to be installed on the single crystal furnace for real-time acquisition and monitoring to ensure stable environmental parameters during the single crystal silicon growth process.The Equal-diameter growth process is the most important and longest time-consuming part of the single crystal silicon growth process.It demands environmental stability during the process of industry production.In this paper,the research on the diagnosis and prediction method of the "Broken Edge" phenomenon of the Equal-diameter growth process of the TDR115P-ZJS single crystal furnace is carried out to reduce the energy consumption and raw material loss in the production process.It has theoretical significance and practical application value.Based on the analysis of relevant literature at home and abroad,this paper completed the following work:Based on the sensor and related process parameters of the process of Czochralski single crystal silicon growth,the time distribution of the"Broken Edge" during the Equal-diameter process.The set of sample points was extracted with the corresponding data was counted.The collection of sample points,and the corresponding data preprocessing which is used to carry out the training of the"Broken Edge" prediction model for the Equal-diameter process at the abnormal sample point detection model combined with the Gaussian Mixture Model and the Logistic Regression,and the Random Forest discriminant model of the Equal-diameter process "Broken Edge",model trained and parameter optimized,.Then verify the performance of the model with field test data;use the optimized data model to conduct online simulation test through field data.Then get the real-time warning of crystal growth record Discrimination and the correct rate achieves a satisfactory result.The main contents of this paper are as follows:The first chapter introduces the "Broken Edge" problem of the Czochralski single crystal silicon growth process and the equal diameter process,as well as the concept and basic flow of data mining.Then by analyzing the limitations of the statistical process quality control(SPC)method,the advantages of data mining in quality control and prediction are highlighted.The research status of data mining technology in the fields of quality control and prediction,and the current state of quality predictive technology related to semiconductor materials are analyzed.The second chapter introduces the various parameters involved in Jingsheng TDR115P-ZJS single crystal furnace.The characteristics and functions of the abnormal sample point detection model based on Gaussian Mixture Model and Logistic Regression and the discriminant model based on Random Forest are analyzed.Finally,the process data monitoring system involved in this paper is introduced,and the overall process and role of the prediction model in the actual production line are analyzed.In the third chapter,the statistical distribution of the equal-diameter duration of the Equal-diameter growth record is used to extract the sample points,and the sample point set is subjected to 0-1 standardization,mutual information coefficient feature selection and principal component analysis.Feature dimension reduction.The fourth chapter introduces the overall process of training the abnormal sample point detection model with the sample point set containing the principal component firstly.Then,the Gaussian Mixture Model training is performed on the sample point set,which mainly includes the parameter initialization of the k-means,the parameter iteration of the EM algorithm,and the selection of the mixed number using the Bayesian information criterion(BIC)coefficient.Then,the training probability vector is trained by Logistic Regression,and the probability that the sample point is "Broken Edge" is obtained.Finally,the performance of the abnormal sample point detection model was evaluated by analyzing the model accuracy.The fifth chapter introduces the overall process of training the Equal-diameter"Broken Edge" discriminant model based on random forest.Firstly,the single CART decision tree model with Equal-diameter "Broken Edge" is trained and optimized,and the best CART decision tree and its tree depth and leaf nodes are obtained.Then the parameter is used as the training hyper-parameter of the Random Forest.The training and optimization are carried out,and the optimized Equal-diameter "Broken Edge"discriminant model based on the Random Forest is obtained.Finally,the accuracy and recall rate of the model are analyzed to evaluate the performance of the above two data models.The sixth chapter firstly uses the predictive model to perform an Equal-diameter"Broken Edge" prediction of all crystal growth records in the test data set,and compares it with the actual results.Then,the alarm advance time and the Equal-diameter duration of the Equal-diameter "Broken Edge" growth record of the test data set are statistically analyzed to verify the feasibility of the model.The seventh chapter summarizes the work of this paper and looks forward to the future research work.