OISF Prediction Of CZ Silicon Single Crystal Based On Data Mining

With the rapid development of microelectronic technology,the quality of IC grade silicon directly affects the yield and performance of semiconductor devices.At present,CZ(Czochralski,CZ)method is mainly used to grow IC grade silicon single crystal.Oxidation Induced Stacking Faults,OISF(OISF),is a process Induced Stacking fault in CZ single crystal,which is influenced by the growth process parameters during the crystal drawing process.OISF can cause the fluctuation of the surrounding band gap energy,and the scattering center can be generated in the body of OISF,which reduces the lifetime of the minority carrier and the mobility of the carrier,leading to the decrease of the electrical performance of the device.As the line width of integrated circuit design has reached the nanometer level,OISF has an increasing influence on the yield of devices.Silicon single crystal manufacturing enterprises need to sample and test each finished silicon single crystal rod,which not only causes a certain waste of raw materials,but also takes a long time to test,increasing the time cost.A number of sensors are set up in the single crystal furnace to monitor the process data in the process of crystal drawing and keep complete historical data.Using data mining technology can dig up the implicit trend from raw data,feature and correlation information,so in this paper,based on CZ number growth process data,using data mining to analyze the number OISF predicted that prompt existence OISF heavily sb-doped silicon rods for testing samples,the number great OISF detection by whole shift for the sampling,Thus,the utilization rate of raw materials and production efficiency are improved.The main research contents of this paper are as follows:Firstly,Silicon single crystal data were obtained and preprocessed.The growth process data of silicon single crystal were integrated with OISF detection results.With the production process parameters as feature vectors and THE OISF detection results as labels,the integrated data set was constructed.Use boxplot method to clean the extreme outliers in the original data and filter industrial noise;The maximum mutual information coefficient is used to remove the redundant parameters in the feature vector to remove the redundant information in the data set.According to the characteristics of the disequilibrium of the data sets,a sampling scheme based on the sliding mean standard with different sampling rates was proposed to extract the characteristic information of parameter changes and reduce the influence of the disequilibrium of the data classes on the subsequent model training.Principal component analysis(PCA)was used to perform feature dimension reduction on the sampled data set to obtain the principal component dimension reduction matrix.The results showed that the first 10 principal components contained 98.5% of the original growth process parameters.Secondly,the prediction of OISF in silicon single crystal is completed based on four optimized classifier models.The PSO algorithm was used to optimize the hyperparameters of the data mining algorithm,and four classifier models including KNN,SVM,BP neural network and XGBoost were constructed and optimized.The principal component after dimensionality reduction was used as the input feature vector of the classifier,and the performance of the sub-model was evaluated by the 10-fold cross-validation method.Since more attention is paid to the recognition ability of OISF defects after the growth of silicon single crystal,accuracy and specificity are selected as the evaluation indexes of the model,and the results of the four optimized classifiers are compared and analyzed.The results show that the four optimized classifiers have the ability to identify OSIF defects.With the increase of the number of principal components,the performance of the optimized classifier is improved gradually.However,the number of principal components used by each classifier to obtain the optimal result is slightly different.Among them,PSO-XGBoost can achieve the optimal prediction effect by using the first six principal components,with the accuracy up to 0.9433 and specificity up to 0.9310.Finally,the scheme is verified in the actual production process.After each single silicon rod is drawn,the data of single crystal growth process are preprocessed and data samples are extracted.The first six principal components after dimensionality reduction were used as input feature vectors,and the OISF of each sample point was predicted by PSO-XGBoost classifier.When the number of samples predicted to have OISF exceeds 50%,it is determined that the test location has OISF defect;otherwise,there is no OISF defect.Finally,the prediction results are verified by manual sampling.The prediction accuracy of the model reached 0.97,and all OISF defects were completely identified.The results show that the OISF defect detection of silicon monocrystallite rod can be changed from sampling for each rod to prediction by model first,and then sampling for the rod with OISF predicted.Therefore,the waste of raw materials can be reduced and the production efficiency can be improved.