Research On Key Control Problems In The Growth Of Electronic Grade Czochralski Silicon Single Crystal

Czochralski(CZ)method is a widely used technology for manufacturing semiconductor crystals.At present,semiconductor manufacturers desire to increase the crystal diameter in order to minimize the unit cost per chip during wafer processing and reduce dislocation density and micro defect level in order to fabricate integrated circuits with smaller line width.However,the CZ silicon single crystal growth process is essentially nonliner,time-varying and with large delay,the quality of crystal is affected by crystal growth speed,environmental temperature,and time lag characteristics.It is difficult to build mathematical model for industrial field and design a good control strategy.Therefore,establishing control model for the CZ process and designing a control strategy that meets the advanced manufacturing objectives of large-diameter,electronic grade silicon single crystal is of great significance.By integrating theory research with experiment research,this thesis mainly focuses on the constant diameter stage in CZ silicon single crystal growth process.Based on the actual measured data in the process of crystal growth,the identification models suitable for industrial real-time control are obtained and optimized by machine learning.The realization of constant pulling rate control structure is studied deeply and some research results are achieved as follows.(1)For the parameter identification of the nonlinear model of heater power and thermal field temperature in seeding stage,the Ant Lion Optimizer(ALO)is introduced and improved by using Levy flight mechanism,elite competition mechanism,search radius continuous shrinkage mechanism and dynamic search mechanism.The test results of standard functions show that the improved optimization algorithm has faster convergence speed and better global searching ability.Finally,the improved algorithm is used to effectively identify the model parameters,which is verified in the crystal pulling experiment.The improved algorithm lays a good foundation for modeling and control optimization in subsequent chapters.(2)For the heat transfer in CZ process,a new T-S fuzzy identification method is proposed to establish the nonlinear model of heater power and thermal field temperature firstly.Then,a deep learning network,namely stack sparse autoencoder(SSAE),is used to study the dynamic characteristics and input-output relationship between thermal field temperature and crystal diameter.This identification method can further increase the identification accuracy of CZ silicon single crystal growth process and provide a more accurate prediction model for the later model predictive control(MPC).(3)The pulling rate is one of the most important control variables for crystal diameter in the existing crystal growth control system,so the nonlinear model between them in "body growth" stage is studied.In this thesis,a novel hybrid deep learning modeling method is proposed by combining the DBN composed of continuous restricted Boltzmann machine,least squares support vector regression(SVR)and the improved ALO algorithm.The proposed hybrid deep learning modeling method takes advantage of DBN to extract system features and SVR algorithm to solve the problems of high dimensionality and local minimum.Compared with other single machine learning method,the experimental results show that the proposed hybrid deep learning modeling method is effective.(4)Frequent fluctuation of pulling rate is harmful to the quality of the crystal in traditional crystal growth control structure,so a constant pulling rate control structure is proposed,in which pulling rate is no longer changed according to the deviation of crystal diameter,and crystal diameter is controlled only by adjusting thermal field temperature.The deep learning identification model of thermal field temperature and crystal diameter obtained above is adopted as the prediction model,and a learning-based generalized MPC algorithm is used to control crystal diameter.The experimental results show that the proposed method has good control effectiveness and production safety compared with the traditional PID method.The above research work not only obtains a more accurate model of heat transfer during the seeding stage and "body growth" stage of CZ silicon single crystal growth process,but also establishs the model of pulling rate and crystal diameter.Based on the model of thermal field temperature and crystal diameter,a learning-based model prediction control of crystal diameter is realized under the constant pulling rate control structure.It provides accurate control models and a safe control method for CZ silicon single crystal growth process.