Research On Virtual Metrology And Process Monitoring To Semiconductor Manufacturing

The semiconductor processes are large,complex and highly nonlinear,which brings great difficulties to related modeling and control.With the development of intelligent instruments and computer techniques,semiconductor manufacturers have accumulated massive production data.How to extract effective information from the semiconductor production data and utilize it have been getting more and more attention.This thesis focuses on data reconciliation and gross error detection(DRGED),virtual metrology(VM)and process monitoring(PM)in semiconductor manufacturing.The corresponding results are summarized as follows:1)For lack of prior knowledge of existing methods to eliminate random errors and gross errors in semiconductor data,this thesis proposes a correntropy estimator based iterative neural network(C-INN)model to achieve data reconciliation and gross errors detection simultaneously.The model takes the measured variables as input and output.C-INN captures process correlations or constraints among the measurements and conducts data reconciliation.Also,the robust correntropy estimator is incorporated to reduce the effect of gross errors.Instead of using the given constraints,this strategy can effectively solve the DRGED problem through the extracted process relationships and the correntropy estimator.2)To address the problem that the traditional global model is difficult to solve the parameters when predicting single-stage key quality variables and difficult to achieve update,this thesis proposed the just-in-time(JIT)approach,along with the variable shrinkage and selection method using the Gaussian process regression(GPR).For a new query sample,the JIT method is conducted to determine the relevant samples to acquire a similar data set to build a local model.By introducing the penalty to GPR,the Penalized GPR(PGPR)is capable to select the essential variables with limited relevant samples.The prediction confidence interval can also be simultaneously generated.3)Being faced with the multi-stage raw data from semiconductor manufacturing,this work presents two multi-stage VM algorithms based on convolutional neural network,in which the nonlinear features are reasonably extracted for predicting key quality variables.The convolutional neural network based multi-stage VM model(CNN-VM)designs novel data arrangement to process multi-stage raw data directly.Meanwhile,the cascade-connected convolving filters are trained together with the regression part to select effective features for the final prediction and improve the accuracy of VM.In addition,the convolutional neural network-gaussian process regression based multi-stage VM model called CNN-GPR is also developed,which replaces the original fully-connected part with a Gaussian process regression and is calibrated by maximizing of the Bayesian posterior density distribution.The CNN-GPR model is also sensitive to the variations of the quality variables.4)Aiming at the monitoring of complex nonlinear processes in semiconductor processes,this thesis proposes a monitoring algorithm based on multi-scale variational autoencoders and generative confrontation networks(MS-VAE-GAN).The wavelet decomposition is adopted to account for the multiscale characteristics.Also,the adversarial networks(GAN)assist variational auto-encoder(VAE)to generate effective latent variable(LV)models to monitor complex nonlinear semiconductor processes.Combined with the critical scale selection,the effective process monitoring statistical features are extracted to achieve accurate detection of abnormal conditions.