Model Identification And Control Of Silicon Single Crystal Diameter For Constant-diameter Growth Stage

Silicon single crystal is one of the basic materials of the semiconductor industry, which plays an important role in IC industry. With the development of very large scale integrated circuits, high-quality and large-diameter requirements are proposed for mono-crystalline silicon.Czochralski technique based on the conventional control structure is the main method of producing mono-crystalline silicon. Under the conventional control structure, the target temperature tracking curve depends on the artificial experience and the parameters of controllers are obtained by repeated experiments. In addition, the process between thermal field temperature and crystal diameter has some characteristics, such as nonlinearity, large time-delay and slow time-varying. In order to avoid unreasonable target temperature tracking curve and controller parameter setting, the most effective strategy is identifying nonlinear dynamic model between thermal field temperature and crystal diameter and adopting the model-based technique to achieve the crystal diameter control. It can improve the quality of the crystal.In this paper, an identification and control scheme of constant pulling velocity is proposed after studying the growth principle of single crystal silicon and the conventional control structure. In this scheme, the time delay of nonlinear dynamic model is determined by the correlation algorithm and identification data from thermal field temperature and crystal diameter of the constant-diameter growth stage. The Lipchitz quotient is used to determine the input and output order of nonlinear model. The enhanced correlation testing algorithm is used to test the identification model and to adjust the model order after the model parameters are obtained. The achievement of parameters identification is based on the dynamic BP neural network and the stacked sparse automatic encoder respectively. Based on the effective model between thermal field temperature and crystal diameter, the simulation experiment of crystal diameter is carried out by using two control algorithms, which are the generalized predictive control algorithm based on the dynamic BP neural network and the stacked sparse automatic encoder prediction model. The simulation results show that the proposed identification scheme can effectively obtain nonlinear dynamic model between the temperature and crystal diameter.The generalized predictive control algorithm introducing stacked sparse automatic encoder can effectively track the set crystal diameter, but the algorithm is poor in real-time correction. The generalized predictive control algorithm based on dynamic BP neural network can track the setting of crystal diameter accurately and correct the prediction model through online learning to adapt to the slow time-varying system.